Documente Academic
Documente Profesional
Documente Cultură
Instructors Manual
OCTOBER 2012
FHWA-HEP-13-002
FHWA-HEP-13-002
Notice
This document is disseminated under the sponsorship of the U.S. Department of Transportation
in the interest of information exchange. The U.S. Government assumes no liability for the use of
the information contained in this document. This report does not constitute a standard,
specification, or regulation.
The U.S. Government does not endorse products or manufacturers. Trademarks or
manufacturers names may appear in this report only because they are considered essential to the
objective of the document.
Table of Contents
Introduction ..................................................................................................................................... 1
Webinar Schedule ........................................................................................................................... 2
Acknowledgements ......................................................................................................................... 3
Webinar Content ............................................................................................................................. 4
Session 1: Executive Perspective ................................................................................................... 5
Session 1 Questions and Answers ............................................................................................. 64
Session 2: Institutional Issues for Managers ................................................................................ 69
Session 2 Questions and Answers ........................................................................................... 153
Session 3: Technical Issues for Managers ................................................................................. 156
Session 3 Questions and Answers ........................................................................................... 237
Session 4: Frameworks and Techniques .................................................................................... 240
Session 4 Questions and Answers ........................................................................................... 334
Session 5: Population Synthesis and Household Evolution ....................................................... 336
Session 5 Questions and Answers ........................................................................................... 434
Session 6: Accessibilities & Treatment of Space ...................................................................... 436
Session 6 Questions and Answers ........................................................................................... 561
Session 7: Long-Term and Mobility Choice Models ................................................................. 563
Session 7 Questions and Answers ........................................................................................... 637
Session 8: Activity Pattern Generation ...................................................................................... 639
Session 8 Questions and Answers ........................................................................................... 720
Session 9: Scheduling & Time-of-Day Choice .......................................................................... 722
Session 9 Questions and Answers ........................................................................................... 812
Session 10: Tour Mode, Primary Destination, Intermediate Stop Location, and Trip Mode .... 815
Session 10 Questions and Answers ......................................................................................... 904
Session 11: Network Integration ................................................................................................ 907
Session 11 Questions and Answers ....................................................................................... 1009
Session 12: Forecasting and Application ................................................................................. 1011
Section 12 Questions and Answers ....................................................................................... 1109
Introduction
This document contains presentation materials from a webinar series on activity-based modeling
held in 2012. The webinar series was sponsored by the Travel Model Improvement Program
(TMIP), which was created to advance the state of the practice of travel modeling by advancing
research and building the technical capabilities of transport agency staff. The overall goal for the
webinar series was to improve the capacity of Metropolitan Planning Organizations (MPOs) to
evaluate and deploy advanced modeling approaches, primarily focused on activity-based travel
demand modeling. The key objectives of the webinar series were as follows:
Address obstacles to the deployment of advanced models by describing the costs and
benefits of advanced models, in relation to the costs and benefits of existing models. Costs
will include staff time, consultant costs, software and hardware needs, and the time it will
take to deliver results. Benefits will include new and improved performance measures, new
planning policies that can be evaluated and improved understandings of travel behavior to
provide explanations of impacts to decision-makers.
Discuss implementation strategies for advanced models that address specific application
needs, incremental deployment of hybrid models, migration from traditional 4-step planning
models, and the resources and expectations needed to manage the development of activitybased models.
A series of twelve webinars were held to address these objectives with three different audiences
in mind: one session for the MPO executive to understand the big picture and the motivation,
two sessions for modeling managers to consider the institutional and technical issues of
developing, maintaining and updating activity-based models, and nine sessions to educate staff
on the principles, frameworks, and techniques to deploy advanced models, as well as options for
implementation.
Advanced Models
The term advanced models can include a wide variety of forecasting methods that are
developed to support transportation planning, including activity-based passenger demand
forecasting models, tour-based and supply chain freight demand forecasting models, land use
forecasting models (integrated with travel models), dynamic traffic assignment models
(integrated with travel demand models), emissions models, and cost-benefit models. Although
the focus of the webinars is on activity-based models, the material was presented in the context
1
of the larger modeling system for freight, land use, traffic, emissions, and cost-benefits so that
practitioners could evaluate their own approach within this context. A key aspect of the webinars
was to provide practical examples of the benefits of activity-based models for addressing new
transportation challenges, such as transport pricing, shifts in demographic trends such as aging
population, travel demand management strategies, and greenhouse gas emissions.
Diversity of Activity-based Models
Activity-based models have been developed by over a dozen MPOs, are being considered for
development by another dozen MPOs, and are in active use for planning applications in at least a
handful of places. There are two prominent frameworks in active use around the U.S. (CTRAMP and DaySim) and at least two others under development at MPOs (AMOS and
CEMDAP) as well as numerous other academic frameworks in the U.S. and abroad. The purpose
of the webinars was not to dwell on specific platforms but to educate participants on the features
in activity-based models and the differences that exist between approaches. The webinars strived
to represent the different frameworks accurately and fairly to present an objective view of the
possible options. The consultant team selected for the project included representation of the
developers of the two prominent frameworks (Parsons Brinckerhoff and John Bowman/Mark
Bradley) and representatives who have used the other two frameworks (Bhargava Sana of
Resource Systems Group for AMOS and Kostas Goulias for CEMDAP). Nearly every webinar
was instructed by a representative of each primary firm (Resource Systems Group and Parsons
Brinckerhoff) to represent the different frameworks and experiences adequately during each
webinar. Material was reviewed by a set of key technical advisors, including John Bowman,
Mark Bradley, and Kostas Goulias, to ensure that all aspects of the different frameworks are
adequately represented.
Webinar Schedule
The webinars were held over eight months in 2012, as shown in Table 1. Also shown are the date
that each webinar was held and the instructors for the webinar. In general, the first instructor
listed was the lead instructor and primarily responsible for content, though in most cases both
instructors and a number of other consultant staff contributed significantly to content as well. As
noted above, the webinar series was presented in two parts; the first three sessions focused on
agency management contemplating moving to an activity-based model for their region, while the
second nine sessions provided more technical detail on the formulation, theory, and mechanics of
activity-based models and their application to a variety of policy scenarios.
Session Description
Numbe
r
Executive and Management Sessions
Date
Instructors
Executive Perspective
February 2
February 23
March 15
2
3
Technical Sessions
Activity-Based Model Framework
April 5
April 26
May 17
June 7
June 28
July 19
August 9
Network Integration
August 30
September
20
4
5
6
7
8
9
10
11
12
Maren
Outwater, Joel
Freedman
John Gliebe,
Rosella Picado
Joel Freedman,
Maren Outwater
John Gliebe,
Joel Freedman
John Gliebe,
Peter Vovsha
Joel Freedman,
Kostas Goulias
Maren
Outwater, Peter
Vovsha
Peter Vovsha,
John Gliebe
Peter Vovsha,
Maren Outwater
Joel Freedman,
John Gliebe
Joe Castiglione,
Peter Vovsha
John Gliebe,
Peter Vovsha
Acknowledgements
This project was sponsored by the Travel Model Improvement Program (TMIP), which was
created to advance the state of the practice of travel modeling by advancing research and
building the technical capabilities of transport agency staff. The TMIP project manager was
3
Sarah Sun. The webinar series was developed and presented by a consultant team which included
Resource Systems Group (RSG) and Parsons Brinckerhoff (PB). John Gliebe served as RSG
project manager, and Joel Freedman was the PB project manager. Stephen Lawe (RSG) also
provided management support for the project. Content was developed and delivered largely by
the following staff: John Gliebe (RSG), Maren Outwater (RSG), Joel Freedman (PB) and Peter
Vovsha (PB). The following staff also provided content and presented material: Rosella Picado
(PB), Joe Castiglione (RSG), Greg Erhardt (PB), Kostas Goulias (University of California Santa Barbara), Bhargava Sana (RSG), Nazneen Ferdous (RSG), and Jason Chen (RSG). John
Bowman, Mark Bradley and Kostas Goulias reviewed and the material and made
recommendations. RSG staff members Bhargava Sana, Brian Grady and Sumit Bindra were
responsible for media production, setting up the webinar software and technical issues.
Webinar Content
The following pages of this document contain the content of each webinar, including the slides
and speaker notes. The questions and answers from the mid-point break and the end of the
webinar are given at the end of each webinar session.
Page 1
Activity-Based Modeling
Session 1: Executive Perspective
February 2, 2012
This is the first of twelve activity-based modeling webinars that we will conduct over the next
nine months. This session is designed as a high-level view of activity-based models, designed for
executives. The next two sessions are designed for modeling managers. The remaining nine
sessions are technical in nature and are designed for modeling staff.
Page 2
Acknowledgments
This presentation was prepared through the collaborative efforts
of Resource Systems Group, Inc. and Parsons Brinckerhoff.
Presenters
Maren Outwater
Joel Freedman
Media Production
Bhargava Sana, Brian Grady
Resource Systems Group and Parsons Brinckerhoff have developed these webinars
collaboratively, and we will be presenting each webinar together.
Maren Outwater and Joel Freedman are co-presenters. They were also primarily
responsible for preparing the material presented in this session.
Stephen Lawe is the session moderator.
Content development was also provided by John Gliebe, Peter Vovsha, and Rosella
Picado.
Bhargava Sana and Brian Grady were responsible for media production, including setting
up and managing the webinar presentation.
Page 3
Learning Outcomes
How travel demand models are used
Benefits and limitations of activity-based models
Why current models cant answer certain policy
questions
Time and resources needed to implement an activitybased modeling system
At the end of this presentation, you should understand the following executive viewpoints on:
Page 4
Outline
(Maren Outwater) I will cover an overview of activity-based models, including providing some
specific practical advantages of their use. In addition, I will cover some of the challenges and
limitations of using activity-based models to provide a balanced perspective (activity-based
modeling is certainly not appropriate for every agency or every purpose). Then, Joel will cover
examples of policy evaluations where activity-based models have an advantage over traditional
methods. Lastly, Joel will discuss the staff and resource requirements of activity-based models.
Page 5
Terminology
Activity-based
model
Tours
Trip-based model
Advanced models
Integrated modeling
system
10
Integrated modeling system involves integrating economic, land use, travel, traffic and air
quality models to provide sensitivity to a broader array of variables. We will not be discussing
integrated modeling systems today, but wanted to provide the context for how activity-based
models are typically used in planning.
11
Page 6
Key Concepts
Activity-based models
provide sensitivities to policies and more intuitive analysis
than existing methods
produce many performance measures that are not possible
with existing methods
do not necessarily take longer or cost more to develop and
apply than existing methods
An all-new activity-based model is a similar level of effort and cost to
developing an all-new trip-based model
An incremental change to an existing activity-based model is similar
in effort and cost to an incremental change in a trip-based model
One of the most important reasons to move to an activity-based model is to provide sensitivities
to policies that are not possible using existing methods. Pricing policies have been pushing many
MPOs into activity-based models because prior models did not have sensitivity to price on
demand, destination or route choice. Another strong benefit is that many performance measures
that are important for decision-making are now possible. For example, traveler benefits accruing
to different populations can be provided to assess the equity of transportation investments.
Now that the first wave of activity-based models have been developed, the time and cost of
developing a new model does not necessarily take longer or cost more. It is difficult, of course,
to make an apples-to-apples comparison of these costs, but some agencies have developed
activity-based models with the same timeframe and costs as a trip-based model.
12
Page 7
Travel demand models have been used in planning to provide information for decision makers.
They do not represent a decision, but allow objectivity in the evaluation of alternatives and the
potential impacts of transportation policies. They can also provide insight on the specific benefits
or limitations of an alternative. Models are also quite useful to better understand the impacts that
various futures, such as changes in gas prices, will have on travel demand. The forecast
assumptions used in travel demand models should be transparent and evaluated through
sensitivity tests to better understand the uncertainty of forecasted input assumptions.
13
Page 8
whether to travel
where to travel to
when to travel
how to travel
Activity-based models are more intuitively correct than traditional models because they closely
follow an individuals decision-making process, whether to make a trip outside the home (or
engage in activities at home), where this activity will take place, and when and how to get there.
Results of activity-based models tend to be more intuitive than trip-based models also. This is
because the modeled relationships underlying in the outcome behavior are more intuitive.
14
Page 9
2-Disc
>75:00
9 P.M.
23
One concept in activity-based models is to model the full daily activity pattern and set schedules
to fit these activities and the travel associated with them into a single day. Typically mandatory
activities, such as work, are scheduled first and discretionary activities, such as shopping or
eating out, are scheduled into remaining time periods.
15
Page 10
Home
Walk
Work
Lunch
Walk
Car
Car
Gas
Station
Daycare
Center
Car
Car
7 trips
2 tours
4 stops
1 stop
Grocery Store
10
Another concept is that trips are part of a larger tour that may accomplish one or more activities
and that all trips on a tour should be linked. For example, if you take your car in the morning to
work, then you must use your car for running errands on the way home. You may also go out to
lunch during the day, which represents another tour. Changes in this system may prompt you to
go home before running errands, which means more trips and possibly different destinations,
modes, or timing for these trips.
16
Page 11
11
Activity-based models are consistent in their representation of travel behavior, which produces
more consistent responses to changes in the transportation system. So, a change to the
transportation system will affect whether someone will make a trip, where they make that trip,
how and when in the same way. Trip-based models do not have the same level of consistency
throughout the process. The other important aspect about activity-based models is that there are
significantly more details and resolution on travelers, space and time, which provides more
information on transportation impacts for decision-making.
17
Page 12
number of persons
housing tenure
residential building size/type
number of persons age 65+
number of persons under age 18
number of persons that are part
of the family
number of children
household income
number of vehicles owned
number of workers
number of students
Person Attributes
relationship to householder
gender
age
grade in school
hours worked per week
worker status
student status
12
For example, activity-based models can take advantage of additional household and person
attributes that are available in trip-based models in a more limited fashion. These include
household attributes and person attributes, which are listed on this slide. Activity-based models
utilize these attributes by synthesizing a population based upon Census data records.
18
Page 13
13
Attributes can also be derived based upon several explicit attributes listed on the previous slide.
One example of an important derived person attribute is that of value of time, which can be
estimated for each person and may also vary by trip purpose or time of day. This additional detail
is necessary to evaluate pricing policies such as HOT lanes, cordon pricing, or tolls by time of
day.
19
Page 14
Activity Purposes
Work
School/College
Personal Business (e.g., Medical)
Shopping
Meals
Social/Recreational
Escort Passenger(s)
Joint Participation
Home (any activity which takes place within the home)
14
Activity-based models typically have many more purposes than trip-based models so that these
can be associated with specific land uses. Often college trips are separated from grade-school
trips, in order to send the right trips, by mode and time-of-day, to the right destination. Escorting
passengers and joint participation in travel provide the means to track the interactions of persons
in a household so that decisions that affect this joint travel are connected. Eating meals is often
modeled as a separate trip purpose from other discretionary travel.
20
Page 15
Activity-Based
Simulation of individual
households and persons
Trips are chainedmodeled as
part of tours, sub-tours and
larger daily activity patterns
Starting and ending time of
activities are modeled choices
Built environment and
accessibility variables affect
travel demand
Market stratification is a
function of individual and
household attributes
15
Many of you have employed trip-based (or 4-step) travel demand forecasting models for
planning purposes at your agencies. I am going to talk about some of the benefits and limitations
of activity-based models in a minute, but wanted to start with a simple comparison of the
approaches.
Most activity-based models simulate individual travel, whereas most trip-based models
generate aggregate zonal estimates of travel;
Most activity-based models model trip timing as a choice, whereas most trip-based
models use fixed factors for trip timing;
Most activity-based models show how accessibility and the built environment affect
travel demand, whereas most trip-based models do not; and
Trip-based models have limited market segmentation capabilities, whereas activity-based
models do not.
21
Page 16
16
One of the best features of activity-based models is that travel choices are based on round trips
and daily activity patterns. For example:
If I need to stay late at work and there is no bus home at that hour, I will not choose to
ride transit to work regardless of how good the service is.
If I decide to run errands near work at lunchtime, then I wont need to stop on the way
home.
If I am telecommuting to work or school, then I wont need to travel at all.
If there are new tolls on the system, I may choose to shop somewhere closer to home or
on-line.
All of these factors are modeled consistently by the behavioral processes in an activity-based
model.
22
Page 17
17
Another important advantage is that the additional detail in the models provides many more
measures of performance for decision-makers. For example:
Travelers benefits can be attributed to different populations, such as low income groups,
to evaluate the equity of specific alternatives;
VMT or emissions outputs can be attributed to households to understand who is causing
these impacts and where they live;
Congestion can be evaluated by half-hour time periods to understand the impact of
pricing policies or capacity investments on delay; and
Traveler benefits can be attributed to clusters of employment that are important for
economic development.
Current trip-based models are not equipped to handle any of the above measures.
23
Page 18
Shopping Trip
Frequency
Time
Period
Work Activity
Arrival/Depar
ture Times
District
Mean Trip
Length
Age Group
Time
Period
Gender
Value of
Time
Mode Share
Income
Group
Trip
Purpose
Mode Share
of Persons
Within mile of
Transit
Parcels
Walk
Trips/Pers
on
Tolls paid
Trip
Purpose
District
TAZ
18
There are many more examples of performance measures that are possible because activity-based
models are based in individuals, which can be summarized across any number of traveler or trip
characteristics. These measures include time spent in various activities, frequency of travel for
various purposes, and person-type summaries of model outputs.
24
Page 19
19
Spatial detail in activity-based models has been developed at the parcel level, the micro-zone
level, or the traditional analysis zone (TAZ) level. The increased detail of parcels and microzones offers more precision, more information for reporting, and more intuitive results. For
example:
The built environment can be represented by buffers of population and employment within a
certain distance of transit stops or parking and by network or urban densities. For example,
transit oriented development can be specifically represented. Non-motorized travel (walk and
bike) and walking to transit also can be explicitly modeled with this additional spatial detail.
25
Page 20
20
Activity-based models are typically much more detailed temporally as well. Often time is
measured in 30 minute time intervals, if not smaller. This provides benefits for evaluation of
operational strategies at the regional level as well as traffic operations at a local level. With this
additional level of detail, analysis of dynamic pricing strategies is possible.
26
Page 21
9%
8%
7%
% of Regional Travel
6%
5%
4%
3%
2%
1%
0%
1
3
EV
8
AM
9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
MD
PM
EV
21
Here is an example of additional temporal resolution in the Jacksonville model. The variations
within a traditional broader time period are significant and may produce misleading results when
an average volume or delay is calculated.
27
Page 22
22
When we synthesize population for the activity-based model, we draw samples of households
with the representative characteristics from the Census. This is one example of a Monte Carlo
simulation approach that is used throughout activity-based models to simulate an individual
characteristic within a known distribution. The simulation can produce different results each time
because of the random draws, but the random draws can also be fixed to produce the same
outcome each time. The range of outcomes can be used to reflect the uncertainty in models of
this type.
28
Page 23
23
The visualization of results in activity-based models is possible because of the additional spatial
and temporal detail and market segmentation that are contained in the models. For example, this
plot of change in real estate prices for each parcel in the Seattle region (1.2 million) shows a
positive change in price due to expanded highway capacity.
29
Page 24
Model features
Optimized software
Hardware
Run time
24
One of the bigger challenges for activity-based models in the past has been the development of
new software platforms, which are now more stable than they were in the beginning. The
computational challenge for these software platforms has been the tradeoff between modeling
features, optimization of the programs, more expensive hardware and run times. Each agency
may identify one or more of these as objectives and must tradeoff the others in order to achieve
the objective. For example, if I want to limit run time, then I will need some combination of
fewer model features, more optimized programs, and more expensive hardware.
30
Page 25
25
While more complexity is possible, it is not always desirable, and it should be tailored to the
region's needs. Tradeoffs for behavioral and spatial realism are inevitable. It is also important to
note that activity-based models can be developed in phases to add detail over time.
31
Page 26
Activity-based models offer an advantage in that many new types of data can be utilized and the
models can take advantage of more detailed data. Activity-based models also can be
implemented with primarily traditional data sources, but this will limit its advantages so
incremental improvements should include enhancements to the data. Activity-based models use
traditional data in more rigorous ways, so the quality and completeness of these data are more
important (and also easier to check and correct).
32
Page 27
33
Page 28
28
Many MPOs that have invested in the development of an activity-based model are motivated by
the need to model pricing policies such as toll lanes, high-occupancy toll lanes, parking pricing,
and/or congestion pricing schemes. Activity-based models are able to represent multiple
dimensions of travel choices that are affected by pricing policies, such as route choice, time-ofday, mode choice, location choice, and parking location choice. Longer-term decisions such as
work and school location choice, vehicle ownership, and transit pass-holding can also be
affected, and those affects can be modeled. Appropriate sensitivities can be represented in the
model by income, household structure, and mobility resources such as auto ownership.
34
Page 29
Example:
Manhattan
Congestion
Pricing
Study
Congestion Pricing
Zone Boundary
Central
Business
District
Congestion Pricing
Zone Portals
29
One of the first activity-based model applications for a major pricing project in the United States
was the application of the New York Metropolitan Region (NYMTC) activity-based models to a
congestion pricing policy for Manhattan. The application tested a number of congestion pricing
schemes, including a cordon pricing scheme, where all auto trips crossing the zone boundaries
indicated on the slide were charged a fee.
35
Page 30
Would require
documentation of
inability to take transit
Level of
Discount
Taxi, Transit
FREE
FLEET
FLEET
Toll-payer Fee-bate
$1 off
50% off
Disabled Drivers
50% off
Zone Residents
50% off
Low-Emission Vehicles
HOV/Carpool
May be accompanied by
investment in Means-Based
Fare Assistance Program
30
Another congestion pricing application involved the San Francisco County Transportation
Authority (SFCTA) activity-based model. This shows an example of one of the toll policies
explored in the study. The complexity of the policy, in terms of the types of discounts offered to
different user groups, is difficult to represent efficiently with a trip-based model.
36
Page 31
0.16
Income
Income
Income
Income
0.14
0.12
0.1
$0-30k
$30-60k
$60-100k
$100k+
0.08
0.06
0.04
0.02
0
$-
$5
$10
$15
$20
$25
$30
31
A key assumption in any road pricing study is travelers value of time, which determines the tolls
that travelers are willing to pay to achieve certain travel time savings. We know from many
surveys and studies that values of time are situational and that they vary greatly, from person to
person and even for any given person, depending of the situation. The SFCTA model represents
this value of time variability explicitly, and doing so helps to obtain a more logical response to
tolls from the model.
37
Page 32
32
Travel demand management schemes are another policy application that activity-based models
are particularly well-suited for. Travel demand management strategies seek to change travel
behavior in order to reduce congestion and improve mobility, and include strategies such as
telecommuting, flexible work schedules, and rideshare programs. Though it is difficult for any
model to predict participation in such programs, it is possible to use a scenario-based approach in
order to model the programs effects on transport demand, congestion, and air quality. A
scenario-based approach involves making assumptions about participation rates (or borrowing
rates from other existing programs) and adjusting model demand to match those assumptions.
The model is then run to determine the impacts of those assumptions.
38
Page 33
Adj/Orig
0.83
1.22
1.12
1.25
1.23
1.01
1.10
0.83
0.93
Original
Adjusted
6
5
% of Tours
Target: Fulltime
Workers
Original
Adjusted
94,408
78,472
115
140
8,070
9,023
13,519
16,848
10,531
12,938
3,817
3,842
13,076
14,360
27,949
23,211
171,485
158,834
Work
School
Escort
Pers Bus
Shop
Meal
Soc/Rec
Workbased
Total
4
3
2
1
15.00
14.00
13.00
12.00
11.00
10.00
9.00
8.00
7.00
6.00
5.00
4.00
3.00
2.00
1.00
Duration
33
For example, a flexible schedule scenario was run using the Burlington, Vermont activity-based
model. The scenario assumed that there would be approximately 20% fewer work and workbased tours as a result, but with longer work tour durations. The tour generation and time-of-day
choice models were adjusted according to these assumptions, and the model was run to
determine the impacts on other dimensions of travel.
39
Page 34
TDM
3000
2000
1000
0
-1000
-2000
-3000
02:00
01:00
00:00
23:00
22:00
21:00
20:00
19:00
18:00
17:00
16:00
15:00
14:00
13:00
12:00
11:00
10:00
09:00
08:00
07:00
06:00
05:00
-4000
04:00
03:00
TDM: Demand
Impacts
4000
TDM-WORK
3000
TDM-NONWORK
2000
1000
0
-1000
-2000
-3000
02:00
01:00
00:00
23:00
22:00
21:00
20:00
19:00
18:00
17:00
16:00
15:00
14:00
13:00
12:00
11:00
10:00
09:00
08:00
07:00
06:00
05:00
04:00
03:00
-4000
34
The results shows a 4% overall reduction in trips, with reduced peak period and midday travel,
but more early AM and evening travel (due to the longer work hours). There were also more
non-work trips in the morning and the evening, as workers seek to fulfill travel needs (such as
shopping and escorting) at other times in the day.
40
Page 35
TDM
800
600
400
200
21:00
22:00
23:00
18:00
19:00
20:00
15:00
16:00
17:00
12:00
13:00
14:00
9:00
10:00
11:00
6:00
7:00
8:00
3:00
4:00
5:00
0:00
1:00
2:00
250
200
150
100
300000
50
BASE
23:00
22:00
21:00
20:00
19:00
18:00
17:00
16:00
15:00
14:00
13:00
12:00
9:00
11:00
10:00
8:00
7:00
6:00
5:00
4:00
3:00
2:00
0
1:00
250000
0:00
TDM
200000
150000
TDM
400
100000
300
200
50000
100
23:00
22:00
21:00
20:00
19:00
18:00
17:00
16:00
15:00
14:00
13:00
12:00
9:00
11:00
10:00
8:00
7:00
6:00
5:00
4:00
3:00
2:00
1:00
0
0:00
23:00
22:00
21:00
20:00
19:00
18:00
17:00
16:00
15:00
14:00
13:00
12:00
11:00
9:00
10:00
8:00
7:00
6:00
5:00
4:00
3:00
2:00
1:00
0:00
35
Only slight declines were observed in vehicle-miles of travel (VMT), with slight increases in the
evening.
41
Page 36
Policies: Transit
Destination and mode choices for round trips (tours) affect
destination and mode choices for individual trips
Tour-level destination and mode choices consider both
outbound and return availability, travel times and costs
Added detail from home to the transit stop and from the
stop to the destination and for local walk and bike travel has
improved accuracy
Transit fare passes and drivers licenses can be explicitly
represented
Built environments affect station area ridership
Activity-Based Modeling: Executive Perspective
36
Activity-based models have also been successfully used for major transit applications, including
New Starts forecasting. Activity-based models offer a number of advantages over trip-based
models for transit analysis. Because activity-based models consider round-trip levels-of-service,
PM peak and evening transit service can affect transit demand throughout the day. Transit fare
policies can be better modeled by explicitly modeling transit fare pass ownership at a personlevel instead of a trip level. Increased spatial accuracy between the origin\destination and the
transit stop results in a more realistic representation of access and egress time.
42
Page 37
37
The New Central Subway was the first New Starts project in the United States to be evaluated
with an activity-based model. This project involved the evaluation of a 1.4-mile long
underground extension to the Third Street light-rail line in San Francisco, connecting the South
of Market area to Chinatown.
43
Page 38
38
This map shows User Benefits provided by the Central Subway compared to a baseline
alternative, specifically for work tours by destination zone. The green zones are winners; that
is, zones that see an overall improvement in mobility due to the subway. The red zones are
losers; zones that see an overall decrease in mobility due to the subway. In this particular
alternative, there are losses in mobility along the existing Embarcadero light-rail line, due to rerouting of trains to the Central Subway corridor, causing an increase in headway and wait time.
44
Page 39
39
The Sacramento Area Council of Governments (SACOG) activity-based model was used to feed
a simulation model developed for Sacramento State University in order to measure demand for a
bus-rapid transit (BRT) project. The activity-based model produces travel demand in 30-minute
intervals. The simulation model disaggregated demand to and from Sacramento State University
to a more refined zone system. Trips driving to and from campus were allocated to one of the
parking lots on campus, and their choice of mode (walk versus transit) between their campus
destination and the parking lot was explicitly modeled.
45
Page 40
Total Spaces
00 30 00 30 00 30 00 30 00 30 00 30 00
5: 6: 8: 9: 11: 12: 14: 15: 17: 18: 20: 21: 23:
Parking constraints
and policies affect
transit ridership
600
500
400
300
200
100
0
23:00
21:30
20:00
18:30
17:00
15:30
14:00
12:30
11:00
9:30
8:00
6:30
BRT Boardings
5:00
Boardings
Time Period
40
The results of the Sacramento State campus area application are shown. The top chart shows how
parking spaces are utilized throughout the day. As parking lots in more desirable locations fill
up, students and faculty must park further from their on-campus destination. As that occurs, BRT
boardings (shown below) increase. BRT boardings are due to the timing of on-campus arrivals
and departures and the use of the BRT line as an intra-campus distribution system (as well as
demand from the nearby light-rail station which the BRT line also serves). Various parking
configurations were tested with the model.
46
Page 41
41
Activity-based models have been used to test policies involving the environment and climate
change. One useful aspect of activity-based models is that vehicle-miles of travel and emissions
calculations can be traced back to the household, since non-home-based trips are modeled as part
of tours. This makes it easier to describe the effects of land-use policy on emissions.
47
Page 42
42
Here is a plot that shows greenhouse gas emissions by residential parcel, from the SACOG
activity-based model. Households residing in more urbanized areas generate relatively less
greenhouse gas emissions than households living in more rural areas, due to relatively smaller
household sizes, shorter trip lengths, and increased use of non-motorized and transit modes.
48
Page 43
Evacuation Modeling:
Persons Not at Home by TAZ and Hour
43
Activity-based models can be used to perform evacuation modeling. This animation shows the
height of each zone based upon the number of persons in that zone who do not live in the zone,
by hour of the day. These are persons who are traveling for work, shopping, and other out-ofhome activities, which is possible because the activity-based model tracks how people are
spending their time throughout the day. This provides an opportunity to model evacuation plans;
the simulation can be stopped for a specific time period and the behavior of each person can be
modeled based upon supplementary survey data.
49
Page 44
44
There are a number of advantages that activity-based models offer to better address land-use
policy. Activity-based models often use a finer spatial system than the zone, so they are able to
provide a more realistic representation of density, mixed-use land-use, and other pedestrian
environment variables.
50
Page 45
45
The Puget Sound Regional Council (PSRC) model was a hybrid model where the land use and
activity pattern generator were micro-simulated. These micro-simulation model steps were then
integrated with a trip-based destination and mode choice model. These examples come from the
activity-based part of the model. These graphs show the results from a sensitivity test where core
urban highway capacity was doubled (i.e. the same networks as the baseline with a doubling of
the lane capacities for the core urban highway facilities (I-5, I-405, I-90, and SR-520) for the
first graph and halved for the second graph). The changes in the parcel prices, along with
changes in the accessibility, filter down through the land use, workplace location choice, and
activity generation models to produce shifts in VMT (8% increase for double capacity; 10%
decrease for half capacity). Some of these shifts come from more trips and some from longer trip
lengths.
51
Page 46
46
This slide shows the changes in population and employment at regional centers in the Puget
Sound Region (Seattle). These are centers for their transportation plan where they have targeted
new growth. Alternatives that support increases in growth in these centers are considered to be
better than alternatives that do not support this growth. MICs are Manufacturing and Industrial
Centers.
The alternatives are combinations of projects with increasing levels of pricing in each (Alt. 1 has
minimal pricing; Alt. 5 is full network system tolling). Alt. 2 has more highway projects than the
others, and Alt. 5 has more transit. The shifts in land use were modest for the alternatives, as
expected.
52
Page 47
47
Activity-based models represent the effects of transport policy on induced demand through their
inclusion of accessibility variables on tour- and stop-generation components.
53
Page 48
48
These graphs show how the effects of transportation improvements on the land use changes that
we just saw also have an impact on induced demand. The activity-based demand model showed
changes to vehicles owned and number of trips made, differentiated by work and non-work
activity types.
54
Page 49
49
There are a number of staff training issues to consider if an agency is contemplating adopting an
activity-based model. Many of the model components have theoretical roots in choice behavior
theory, so knowledge of discrete choice modeling is essential. In addition, the model system
application may require more custom scripting and programming than trip-based models. These
skills are necessary in order to maintain and enhance the system, but may not be necessary to run
the models. Since activity-based models produce databases containing the travel choices of the
synthetic population, it is important to have familiarity with statistical and/or database software.
There are also implications for the development of input data and the maintenance and coding of
networks, depending upon the details of how the system represents space and time.
55
Page 50
50
56
Page 51
51
Model run times depend on several factors, the most important of which is the number of agents
in the model. Models for larger regions, such as the San Francisco Bay or Atlanta regions
typically distribute computational burden across multiple computers because the simulations are
for millions of people. Other issues that may require more computing power include the number
of alternatives in various models, extent of shadow pricing and feedback loops, type of sampling
used for models with large numbers of alternatives, number of time periods and modes skimmed,
and efficiency of program code. Another option for sharing resources is cloud computing, but
documentation is limited (less extensive than for off-the-shelf software) and support must be
negotiated.
57
Page 52
Visitors
Airports
Universities
Commercial travel
Internal\External and through-travel
Other long-distance travel
Special events
52
Just as with four-step models, special market models may be required in addition to the core
resident activity-based model. These markets might include visitors, airports, internal-external
travel, and other markets. These models can either be adopted from existing trip-based methods,
or developed specifically to be consistent with the activity-based model. Tour-based treatments
for many of these markets were recently developed specifically for the San Diego activity-based
model system.
58
Page 53
53
Activity-based models rely upon random number sequences to determine results. Therefore there
is random variation within and across forecasts. In such cases, it is useful to analyze a
distribution of results; particularly for model outputs in which a limited number of decisionmakers are affected (such as a local street volume, or ridership on a low volume transit route).
Such distributions are useful in order to communicate the uncertainty associated with particular
outputs. An alternative would be to fix random number seeds in order to ensure consistent results
across model runs, though it should be recognized that such methods result in only one
realization or outcome from a distribution and could be misleading. A better approach is to
average multiple runs. In all cases, it is important to conduct reasonableness checks and
sensitivity checks on models in order to ensure that models react reasonably to changes to inputs
and are ready to be used for forecasting policies of interest.
59
Page 54
Train staff
Identify a champion
Activity-Based Modeling: Executive Perspective
54
We recommend developing a data collection and model development plan prior to embarking on
an activity-based model development project. This helps plan for funding and keeps the overall
project on track. Some initial decisions to be made are the extent of new data collection and
whether to develop the model all at once or in phases through a number of years. One should
plan on thorough calibration, validation, sensitivity testing, training and documentation.
Understand that certain new features may involve some risk, at least to schedule, as research and
development takes time. In addition, it is helpful to have an activity-based model lead or
champion at the agency to keep staff briefed on model development and application activities
and to secure funding.
60
Page 55
Further Research
Advancements in modeling decisions across multiple
dimensions (destination, mode, tours, trips, schedules)
Testing models with information technology policy
parameters
Integration with dynamic traffic assignment models
Transferability of activity-based models
Visualizing and communicating model outputs for
decision making
55
There are many advancements being made in activity-based modeling, some of which are listed
on this slide. They include advancements in discrete choice models related to modeling many
alternatives and multiple dimensions simultaneously, integration with dynamic traffic assignment
models, the transferability of activity-based models, and software and techniques to mine and
visualize the data produced by activity-based models.
61
Page 56
62
Page 57
February 2
February 23
March 15
March 22
April 5
April 26
May 17
June 7
June 28
July 19
August 9
August 30
57
Thank you for joining us this week. The next webinar will be held in three weeks, and will cover
institutional topics for managers.
63
population won't generate as many university students or visitors. If there are differences in
transport supply, that can be reflected as well. However, you need some data for these things.
Maren: Part of this is affected by special events, and there are some special events models out
there. Most are based on trip-based models, but they can be based on activity-based concepts as
well.
Is it possible to incorporate trips coming from outside the area, i.e., the Super Bowl?
Maren: Trips coming from outside the area are generally modeled as external trips or visitor
trips. External trips that come from outside just visiting for the day would be modeled similar to
how they are currently modeled in trip-based system: would enter at cordon location and then are
attracted to special event location or wherever else they are going. Something like a Super Bowl
would have a big impact on both non-residents who travel there for event as well as residents
who attend event. Most special events modeling has focused on evacuation modeling.
Where are agencies getting land use data from and can you define micro-zone?
Joel: Some concepts will be covered in larger detail in later sessions. For micro-zones, imagine
Census blocks. It's not quite as small as a parcel, but smaller than a typical zone. Land use data
comes from a lot of sources. Census data describes households by number, income, size,
workers, age, and gender distributions. These can all be used to control the attributes of the
synthetic population. State agencies frequently have data available as well. Employment data
comes from the Unemployment Division and private data vendors for the base year. For future
years, one is left to the same methods used for trip based model allocation or using some version
of a land use model. There are agencies out there that develop land use models that work with
the activity based model. A land use model would provide household and job information to the
activity based model for future years.
Maren: One of the sources we use for land use at the parcel level is the tax assessor's data. These
exist everywhere but the quality can vary. It has spatial information and buildings that are on a
parcel.
Does ABM also build an OD matrix or something to show individual movement and assign to
network in a conventional way?
Joel: Yes. As mentioned, the end result looks like a travel survey for every person in model. That
data can be aggregated any way the user wants to aggregate it. It is aggregated into zones for
assignment. There is also research to integrate models with dynamic traffic assignment software.
After a model is run, some scripts are written to summarize results by origin, destination, time
period, and mode, and those matrices are assigned to network using traditional travel demand
modeling software. Results are fed back into model, and the model is run iteratively.
65
Please further address issues of using activity based models for new starts. The FTA's Summit
program requires zone-to-zone trip tables for transit share. It seems like using a disaggregate
approach could confuse comparison of transit for scenarios.
Joel: Good question. Disaggregate choices happen in activity based models, and Summit wants a
fixed set of inputs from mode choice program. What's required here is that certain decisions are
turned off in an activity based model to provide the inputs for the Summit program. For example,
tour generation, synthetic population, destinations, time of day held constant. What you allow to
vary is the mode choice component. You then save the probabilities and utilities of individual
decision makers for modes and then aggregate in a mathematically sound way that yields the
overall probability that Summit wants. You are saving individual data and then aggregating for
summit input files. You're essentially turning off sensitivities in model. It's important to hold
random number seed constant so that the changes aren't influenced by Monte Carlo simulation.
Is there any reason to believe that UB computed from a 4-step model would be higher or lower
than an activity based model?
Joel: They will certainly be different. The FTA process assumes non-home-based (NHB) trips
can switch modes depending on changes in probability, whereas an activity based model links all
trips on tours. So NHB trips are constrained by the mode choices made for the tour by traveler.
There needs to be some way to account for this difference to provide a level playing field with a
trip-based model.
Is Monte Carlo run in the model or as a post-processor?
Maren: Monte Carlo is an integral part of the process. It's used a number of times. One of the
most obvious places is the population synthesizer at the beginning of model process. It does use
random numbers, so either fix seeds in order to get same outcome each time, or allow random
number generator to create a randomness representing uncertainty. Monte Carlo is embedded in
methodology for a number of the models.
Normally, do you loop back to distribution or generation phase? How does feedback work?
Maren: Feedback in activity based modeling goes back to top step of the process. Every
component incorporates variables that represent supply side or accessibility. For example,
residential choice and employment choice are affected by accessibility, as well as auto
availability and travel choices. A number of components are affected by changes in accessibility.
These aspects are not incorporated into 4-step models because a change in accessibility would
just be reflected in the distribution.
You mentioned averaging. Is convergence an aspect of activity based modeling?
Joel: Yes. When we're feeding travel time skims back into model, we check for convergence the
same way as in a 4-step model, i.e., changes in volume on links. We make sure we've run the
66
model enough times to achieve stability. We do have some guidelines on the Monte Carlo
process for how many times it's necessary to run in order to get expected values.
In the case of averaging results, is it possible or practical to compare different scenarios?
Joel: Yes, but it depends on which scenarios you are looking at. Work that's been done to date
shows that models respond appropriately. Some applications don't require averaging, and seeds
change very little. It is very possible to compare two scenarios with one model run for each
policy. But again, if you're looking at a policy with less decision makers exposed to the policy,
you need to average. It depends on the policy, but for most policies that have been tested the
Monte Carlo process has not impeded the ability of agency to compare scenarios.
Maren: This is specific to decision makers. Some would prefer not to see a range of outcome and
would prefer to see a single outcome. Others want to see the range. It comes down to a
preference of the policy boards.
What visualization tools do you typically use?
Maren: Visualization tools can vary. Some are the same as a trip-based model. Mapping is
usually done with GIS. A variety of statistical and numerical summaries may be made of the
segmentation data. Those summaries are different from a trip-based model due to the variety of
data.
Joel: There is some custom software to do quick database summaries for scenarios and
differences. It depends on familiarity of staff with different procedures.
Can you give an example of a successful transferred activity based model?
Joel: Lake Tahoe transferred their model from Columbus, Ohio. The budget was fairly small and
the model transferred quickly. It was applied to residents, not visitors. A more recent, Chicago
region transfer was hybrid of a couple different models.
Maren: We transferred a model for Jacksonville and Burlington from Sacramento, and applied,
calibrated, and integrated the activity based model with dynamic traffic assignment in the new
areas.
How much do activity based models cost?
Joel: That depends on the region and amount of innovation versus transferring, versus adapting.
It also depends also on the complexity of region. Chicago spent $250K on an initial model to get
the model transferred, and do some estimation and calibration. Subsequently they are spending
more to enhance the model. Tahoe spent about the same amount to transfer the Columbus model
and develop a visitor model. Other regions are spending a million to 1.2 million for model
development. It really depends on the complexity and amount of original model development
versus adaptation.
67
Maren: There might be two different agencies who both want to spend a half a million, but what
they choose to focus on could be entirely different. There is a wide variety of objectives.
68
69
Page 1
Activity-Based Modeling
Session 2: Institutional Issues for Managers
70
Page 2
Acknowledgments
This presentation was prepared through the collaborative efforts
of Resource Systems Group, Inc. and Parsons Brinckerhoff.
Resource Systems Group and Parsons Brinckerhoff have developed these webinars
collaboratively and we will be presenting each webinar together.
John Gliebe and Rosella Picado are co-presenters. They were primarily responsible for
content, along with Joel Freedman.
Stephen Lawe is the session moderator.
Content development was also provided by Peter Vovsha.
Bhargava Sana and Brian Grady were responsible for media production, including setting
up and managing the webinar presentation.
71
Page 3
February 2
February 23
March 15
April 5
April 26
May 17
June 7
June 28
July 19
August 9
August 30
September 20
3
For your reference, here is a list of all of the webinars topics and dates that have been planned.
As you can see, we will be presenting a different webinar every three weeks. Three weeks ago,
we attempted to provide a somewhat high-level executive view of activity-based modeling.
Today, we will be covering the second in topic in the seriesInstitutional Topics for Managers.
Our objective is to get into a bit more depth on the issues that we have found to be important to
the people we have talked to in our work in activity-based model development. Today we will be
talking about what it takes to transition between a trip-based model operation and one that relies
primarily on an activity based model. We will be talking about development time and costs,
resource allocation, and issues related to productivity.
So, in this webinar we will try to stay away from the more technical issues surrounding activitybased modeling. As you can see by the schedule, there will be plenty of technical detail in the
remainder of the series.
72
Page 4
Learning Outcomes
Typical motivations and concerns of agencies
considering an activity-based model
Familiarity with the evolution of activity-based models
in the U.S.
Development options for migrating from 4-step to
activity-based models
Resources needed to implement an activity-based
model program
Experience with stakeholder acceptance and use
Activity-Based Modeling: Management Institutional
Our audience today is composed of modelers from public agencies, consulting firms and
academic institutions. We also know that there are managers of various levels among you. Our
goal in this webinar is to provide you with more of the institutional context for how travel
demand modeling has evolved to the point where we are today in which there seems to be a
growing demand for more advanced modeling tools. Accordingly, at the end of this webinar you
should have a good understanding of the motivations and concerns that public agencies have
when contemplating moving to an activity based modeling system. To begin to address some of
those concerns, it is helpful to review how activity-based models have evolved over the last
decade or so in different parts of the U.S. To make things a little more concrete, well discuss the
various options that some agencies have followed in developing their activity-based modeling
systems. Resource requirements are always an important issue, and we will share with you some
examples of what some agencies have invested in consulting fees, data development, hardware
and software, and staff resources. Finally, we will discuss some of the experiences to date of
users of these systems, including project use and potential use by stakeholders.
73
Page 5
Terminology
Upfront model development
Phased model development
Transferred model development
74
and validation using local data. This allows the agency to get started fast. We will discuss these
three strategies in more detail later in the webinar.
Page 6
Developed in 1950s
4-step process
Limited by data availability and computing power
Primary applications were planning for highway
capacity--emphasis on vehicle trips and flows
Reliance on simplified trip-based approach
Aggregate relationships
In order to provide context for our discussion, lets step back in time and review how we got
here. Travel demand models were first used in the U.S. during an era in which the Interstate
Highway System was being planned. It was an era of suburban expansion and a post-war baby
boom. Consequently, the focus on modeling efforts in those days was highway capacity
planning. Needless to say, computing power was not nearly what is today, so the process that
was developed, which became the UTMS, was necessarily simple. It was based on the prediction
of aggregate trips being generated from zones, composed of aggregations of households and
businesses, distributed between zones, and assigned to a network to determine how well the
network would perform.
Some of the difficult questions that transportation planners face todaygreenhouse gas
emissions, travel demand management, congestion pricing, transit-oriented development and
75
environmental justicehad not yet emerged as important topics in the early days of travel
demand modeling.
76
Page 7
Trip Distribution
Mode Choice
Network Assignment
The trip-based models of today are really just advanced version of the UTMS process. Here you
see what many of us know as the familiar 4-step process, consisting of trip generation,
distribution and mode choice. Over time, the profession has added explicit representation of
transit and, in some places, pedestrian and bicycle travel modes. With the introduction of discrete
choice models to the profession, models based on utility theory and estimated from individual
observations were an early improvement, although in the end they are still applied to
aggregations of trips rather than to individual travelers. In addition, trips are assigned to
networks that typically represent peak and off-peak travel periods, which provide some
differentiation between level-of-service conditions during different parts of the day.
Another major improvement is the feedback loop in which travel times and costs are fed back
turned into skims tables and fed back into trip distribution and mode choice. This has long been
standard practice in the U.S. It is interesting and relevant to point out here that feedback loops
were mandated as the result of legal challenges and became a recommended best practice for
consistency for air quality modeling. When a particular interest group opposes a proposed action
77
based on a forecast, they challenge the methods used to produce the forecast. In the case of
feedback loops, critics pointed to the need for consistency between the travel times being
produced by the network assignment process and the representation of travel times and costs
being input to the demand models.
78
Page 8
Non-HB
on foot
HB Work
8:00 A.M.
12:00 .P.M.
7:30 A.M.
12:10.P.M.
12:10
home
in zone X
work place
in zone Y
(work)
6:30 P.M.
1:00 P.M.
5:00 P.M.
HB Shop
by auto
restaurant
in zone W
(lunch)
12:50 P.M.
on foot
6:00 P.M.
5:30 P.M.
by auto
Non-HB
Non-HB
grocery store
in zone V
(shopping)
Lets consider how people really travel. Here weve depicted an individual who goes to work at
7:30 a.m., arriving at 8. Around 12 noon, this person walks to lunch and then returns to her work
place at 1. She leaves work at 5 p.m. and stops at the grocery store before going home.
The way this would typically be represented in the trip-based modeling world would be the
following. (Step through HB work, HB shop, and three Non-HB trips). The HB-Work and HBShop trips are in the AM and PM Peak periods. One of the Non-HB trips is in the PM Peak, and
two Non-HB trips are in the off-peak period. We know their modes and trip lengths.
One question that transportation planners typically struggle with is how to explain to stake
holders in your area the impact of particular project, plan or policy on non-home-based trips?
What does a non-home-based trip mean to them? A trip-based model assumes that all of these
trips are independent of one another. It does not account for the fact that all of these trips are
actually part of one large daily activity pattern, anchored around a mandatory work activity. A
trip-based model does not account for the fact that trips are chained into tours and that there is
actually a work-based sub-tour within the larger tour.
79
It also does not account for the fact that, because this person walked to lunch, they do not have
their car available to get back to the office. Further, a trip-based model would not recognize that
this person needed to arrive at work at 8 a.m. and therefore, did not have the time to drive her
son to school since his school is in the opposite direction. So, he has to take the bus. Nor would a
trip-based model recognize that this worker needed the car for work on this particular day
because her planned agenda included a big grocery shop after work. The trip-based model would
also not recognize that persons who work in this location are likely to go out for lunch more
today than ten years ago, because there are now more dining opportunities within walking
distance of this office. An activity-based model would take into account all of this additional
information.
80
Page 9
All of the additional information that an activity-based model takes into account are important,
because in real life trips are not independent from one another and people do not respond to
changes in transportation system level of service changes or policies as if they were. In real life,
trips are organized into tours that make them interdependent. People plan activities at the end of
the day that cause them to make certain travel decisions at the beginning of the day. Mode
choices may be somewhat constrained by household linkages and obligations, such as taking care
of children. The opportunities presented by surrounding land uses may induce people to make
more or fewer discretionary stops. And in the long-run, people do make choices of where to live,
work, go to school, and whether and what types of vehicles to own that are at least partially
based on the transportation environment.
From a technical perspective, this comes down to accurately representing the actual alternatives
available to people in their activity-travel choices. What is really in their choice set? What are
their real short- and long-term elasticities? We will cover the finer points of choice sets and
elasticities in future webinars.
81
Page 10
Road pricing
Variable time-of-day tolls (congestion/time of day)
Area pricing
HOT/HOV lanes
10
Instead, lets talk about policies. How can we better estimate peoples response to changes in
travel costs? For example, how can we better estimate change in VMT as a function of gasoline
prices? If gas prices this summer reach a new all-time high in the U.S., will people take more
transit? Travel less frequently? Make shorter trips? Car pool? Buy more fuel efficient cars? or
forego family vacations and eating out? If high prices persist, will some people choose to work
closer to their residences? These are all legitimate responses that we observe in data, or at least
anecdotally.
These same set of responses are relevant for other policy examples, too. This slide also lists a
number of policies related to how people value their time when faced with changes in travel
costsroad pricing, transit fares, environmental justice. Trip-based models typically do not do a
good job of capturing the multi-faceted response of real people, because the basic unit of analysis
is the individual trip. Important contextual information is simply not there. In addition, trip-based
model make aggregate-level predictions for households of a certain type, but are unable to
82
distinguish between individuals within households. Consequently, they tend to do a poor job of
portraying how individuals value their time.
Page 11
Policies that involve coordination between individuals and timesensitive scheduling constraints
Demographic changes
Household size and composition
Planning to support aging populations
Carpool/shared-ride arrangements
Parking
Capacity constraints/restrictions
11
In addition, because activity-based models represent individual decision makers and operate at a
finer temporal resolution, they are better suited for analyzing policies that require coordination
between individuals and time-sensitive scheduling of activities. As such, they provide more
information for the analysis of policies related to demographic changes, travel demand
management and related commuting options, and time-sensitive issues related to the availability
of parking at different times of day.
In reality, many policies involve both time and scheduling trade-offs along with user willingness
to pay to obtain better level of service. The technical nuances of these trade-offs are something
that we will explore in more detail in future webinars.
83
84
Page 12
Highway Capacity
Project
4-Step Planning
Model
Highway
Performance
More sophisticated models are needed for more complex problems, e.g.
Congestion Pricing
Policy
Activity-based
Model
Traffic and
Revenue from
Tolling
12
Are activity-based models always the best tool for the job? Not necessarily. For problems that
are relatively narrowly defined and for which the likely range of transportation system user
responses is expected to be limited, then a good trip-based model is probably just fine. Deciding
whether and how much to extend roadway capacity in a corridor might be one example.
On the other hand, if you are charged with analyzing policies or plans that involve somewhat
complex policies in which user may be considering multiple options for trading off time and
money, a tool that allows you to model that range of responses is probably what is needed.
85
Page 13
13
What are the consequences of not using the best tool for the job? In a word, it comes down to
credibility. It is important to use a model that is appropriately sensitive. Your trip-based model
might be producing the an appropriate aggregate response, but you may have trouble explaining
it and may be unable to show the impacts on individuals.
As we discussed before, in this business, there are frequent challenges to the credibility of the
methods used to make forecastsespecially when large investments are at stake, or controversial
policies. So that may be a concern.
Finally, although it may be unfair in many cases, there is always the risk that some parties may
perceive your agency as not using the best tools available and therefore a little behind the times.
86
Page 14
Uncertainties of Implementing
Activity-Based Models
Cost
Can it be developed affordably?
Can we afford to maintain it?
Resources
Will it require special technical skills that are difficult to find
in-house?
How long will it take to develop?
Will it have a negative impact on agency productivity (longer
run times, more maintenance, diverted resources)?
14
With all of these reasons to consider moving to an activity-based modeling system, there remains
a great deal of uncertainty as to what it might take to implement such a system. The most
important objective of this presentation is to remove some of this uncertainty.
Typically, people are most interested in how much it will cost, not only to develop but also to
maintain. They often express concern on the effect if will have on staff resourceswhether they
will have the right skill set in-house to run the model or become dependent on a consultant. They
want to know long will it take to develop an activity-based model, and how will it affect agency
productivity.
87
Page 15
Uncertainties of Implementing
Activity-Based Models
Data
Will it require additional data collection?
Household diary surveys
Detailed land use/parcel level data
Additional traffic counts, boardings, etc. for calibration and
validation
Parking supply data
Socio-economic data
15
Our models are only as good as the data used to develop them. Given all the information that
activity-based models are supposed to provide, people naturally expect that a good deal more
data must be required. In fact, much of the data required is similar to what has been collected for
trip-based models, though the level of detail might be greater as activity-based models may be
more sensitive to the accuracy of inputs. In household surveys, activity-based models use more
of the information that is already in the survey diaries.
88
Page 16
Uncertainties of Implementing
Activity-Based Models
Quality
Will it have the desired sensitivity to justify the investment?
Will the methods used in an activity-based model be accepted
in tightly regulated modeling contexts:
EPA conformity, FTA New Starts, NEPA alternatives analysis, LRP?
Will the agency still need to maintain a separate trip-based model?
16
A lot of people also wonder an activity-based model is really as good as advertised. Will the
performance justify the investment?
Most agencies have many years of institutional knowledge in honing their trip-based modeling
skills and refining the models for a variety of important work products, such as conformity
analysis, New Starts applications, NEPA studies, and long-range planning. Understandably,
theyd like some confidence that an activity-based model will perform well under these tightly
regulated scenarios. Naturally, they also wonder if they will need to maintain a trip-based model
as a backup and, if so, can they afford to maintain two model systems.
89
Page 17
Uncertainties of Implementing
Activity-Based Models
User Experience
What is the learning curve?
Will the application software be user friendly?
Will it be comprehensible and easy to explain to
stakeholders?
Can the detailed output of an activity-based model be
transformed into transparent and concise decision-supporting
formats?
Will constituent agencies and consultants be able to use it?
Transit agencies, DOT partners, municipalities, local consultants
Activity-Based Modeling: Management Institutional
17
In addition, there is the user experience on a day-to-day level. What is the learning curve? Will it
be easy to use? Will we be able to explain the model and its outputs to stakeholders?
For modelers who are used to seeing trip-level outputs, such HB-work trips, HB-Other trips and
Non-HB-trips, there is uncertainty in not knowing what the output be like when travel behavior
is expressed in terms of activities and tours. If there is so much more information and output,
will this require sophisticated data mining skills? What software tools have been developed to
help?
Finally, DOTs and MPOs often serve the modeling needs of constituencies composed of transit
agencies and municipalities. In addition, there are often local engineering consultants who have
used their models for years in traffic impact studies and similar work. Will they be able to use
the new activity-based model, or will they insist on sticking with the trip-based methods which
with they are most familiar?
90
Page 18
Burlington
Oregon
Chicago
Ohio
Columbus
Atlanta
Houston
In application
New York
Philadelphia
Jacksonville
Tampa
Miami
Under development
18
The answers to these questions may be found by talking to modelers in regions that have already
adopted activity-based modeling systems. This map shows locations in the U.S. where activitybased models have been developed. As the map key indicates, red dots are locations where a
model system has been finished and is known or thought to be in operation. In some cases, these
might be quite recent. The green dots indicate locations where model systems are now under
development. Interestingly, there are two states shown here in blueOregon and Ohiothat
developed activity or tour-based components for their statewide modeling systems. It is also
interesting to note that all of these systems have been developed within the past 12 years.
91
Page 19
19
To give you some statistics, of the 35 largest MPOs in the U.S., 17 have already or are in the
process of developing an activity-based modeling system. In the last 5 years, all of the large
model development projects have been a move towards an activity-based modeling platform. In
addition, at least three states have decided to encourage the development of activity based
modeling systems for the larger MPOs within their states. This includes Ohio, where the success
of the Columbus model has inspired confidence in its transferability to other large cities in the
state. In California, SB 375 introduced sweeping changes in the way transport planning agencies
analyze transportation and land development, mandating the development of activity-base
models for the largest MPOs in the state. Florida is another recent convert, in which FDOT has
begun to support development of activity-based models in its larger metro areas.
92
Page 20
20
Here is a list of the known activity-based modeling projects that have been implemented in the
U.S. The development of the early pioneersSFCTA and NYMTC actually started in the 1990s,
some of these projects have long development histories, while others benefitted from the work
done on the early models. One example of this was the MORPC model, developed for
Columbus, which provided the basic model structure and software for the Lake Tahoe model. As
you can see, some of these model systems were only recently completed.
93
Page 21
21
Here is a list of some of the known activity-based modeling projects now underway. As you can
see, the number of new activity-based model development projects that have been started within
the past 3 to 4 years is about the same as the number of projects that were completed between
2000-2010. So, the pace of development is accelerating.
94
Page 22
22
You might be wondering what the features are of these various activity-based modeling systems
and how they differ from each other. The design features of the various modeling systems will be
covered in exquisite detail in future webinars. For today, however, it is sufficient to consider the
common features that have become fairly standard across the various at a somewhat high level.
Interestingly, some standardization of model components has occurred, although variable
specifications and certain structural elements differ quite a bit between systems. This slide shows
the common feature of activity-based models in use today in the U.S. These features include:
Page 23
Tour time-of-day
Entire-tour mode
Stop frequency
Stop location
Accessibility of potential
activity sites along the
route between the primary
destination and home
Trip mode
Trip departure time
Activity-Based Modeling: Management Institutional
23
A central concept in these models is the notion that travel episodes are organized into tours and
that people first make decisions about destinations, modes and timing in consideration of the
entire tour, based on expectations of what they want to do. Trip-level decisions are conditional
upon tour-level choices. It should also be noted that this diagram shows a particular sequence of
decisions (destinations, time of day and mode) at the tour level, and at the trip level (stop
frequency, location, mode and time of day). These sequences may vary from one modeling
system to the next and sometimes even within the same modeling system for different contexts.
The important takeaway is that tour-level choices condition trip-level choices.
The other important piece of information is found on the right side of the graph. The accessibility
values of potential downstream choices are fed back up the model chain and used as predictors of
upstream choices. This type of vertical integrity is another comment feature of activity-based
models, at least the ones most commonly used in practice.
96
Page 24
24
Trends in activity-based models have evolved from an initial set of models that implemented
activity and tour-based concepts that we now consider to be fundamental to activity-based
modeling platforms. Over the years, developers of these models have refined their designs in
different ways. Some have added spatial detail to support the analysis of land use impacts on
transportation accessibility, its effects on discretionary stop making, and to provide higher
resolution analysis of pedestrian movements. Other model developers have focused on the
dynamics within households and have explicitly modeled some of the sub-decisions that people
make when coordinating drop-off and pick-up arrangements, scheduling around children, and
joint activity participation. Even more recently, model developers have begun to move toward
finer temporal resolution in order to better reflect time sensitivity to changes in travel costs over
the course of the day.
97
Page 25
DaySim
CT-RAMP
25
This slide shows recent flow diagrams for the two most common activity-based modeling
frameworks in the U.S. In recent, years these systems have been branded with the names
DaySim (shown on the left) and CT-RAMP (shown on the right). We wont cover the details
of each model system todaythat will be explored in subsequent webinars. I put them side by
side to point out how these two modeling systems are becoming increasingly similar. Both
systems share the common features we just discussed, including a few items, such as free
parking eligibility, that are non-standard, but necessary for analyzing TDM policies.
DaySim (left) was first implemented in Sacramento utilizing a very detailed parcel-based
representation of land use, which was a departure from predominately TAZ-based systems. The
earliest implementation of the modes now known as CT-RAMP (right), the Columbus, Ohio
model, focused on explicit modeling of interactions between household members. This came at a
time when other activity-based modeling systems were modeling individual activity patterns,
with interaction between household members more of a correlated attribute rather than a hard
constraint. In a recent specification in Seattle, however, we now see joint activity generation and
98
scheduling being added to the DaySim model. And in a recently completed version of CTRAMP in San Diego, a more detailed spatial unit of analysis was implemented called microzones (though not shown in this diagram). To be sure there are differences in the fine details of
model structures and specifications, but it does appear that a common vision of functionality has
begun to emerge.
99
Page 26
26
It is also important to highlight what activity-based models have in common with trip-based
models. To start, agencies implemented activity-based models should expect to use the same
commercial travel demand modeling packages for network assignment and scenario
management. The one difference might be using more highway and transit assignment periods in
order to take advantage of the temporal resolution reflected in the activity-based model outputs.
Socio-economic inputs and land use data should be quite similar, but some activity-based
modeling systems require that they be maintained a more disaggregate spatial resolution. Once
again, we are talking about a parcel-based system or micro-zones.
Activity-based models of the kind we are covering here are focused on resident travel. Public
agencies considering the move towards an activity-base modeling system should be prepared to
maintain truck and other commercial vehicle movement models, just as they did for their tripbased model. Similarly, Internal-External trips will be need to be generated and distributed
through a separate process, probably the same one used for the trip-base model. Likewise, airport
and visitor trips are not typically covered in an activity-based modeling system for an urban area.
100
While it is possible to develop activity-based commercial travel models and even tour-based
visitor models, those represent completely separate models and processes.
101
Page 27
27
Given the set of available activity-based model features and how they might interact with other
system components, what are the options for an agency interested in transitioning from a tripbased model system to an activity-based model system? Can one innovate incrementally by
gradually replacing trip-based model components with tour-based model components? Or might
it be possible to add sensitivity to an existing trip-based modeling system so that it behaves more
similarly to an activity-based modeling system? Some agencies have experimented with adding
components to their trip-based modeling systems to do just that. One strategy is of course to add
more market segmentseither in terms of trip purposes, or socioeconomic segments at an
attempt to explain more variation. One example might be to make assumptions on a certain
percentage of trips being affected by travel demand management policies based on historic
participation rates. Another example would be the 4-D-plus post-processing tool, which was
tested out by SACOG prior to their usage of their activity-base model. The 4-D process (density,
diversity, design and destinations) was intended to adjust trip generation rates according to these
four dimensions of land use.
102
At some point, however, it may become apparent that adding these features costs time and
money. Moreover, it may produce a complicated modeling system that does not do everything
that an activity-based model could do, has lengthy run times, and requires excessive storage of
trip tables and skim matrices.
103
Page 28
Phased development
Multi-stage effort, replace 4-step model components
gradually, multiple RFPs
28
This slide lists three generalized approaches for the development of activity-based modeling
systems. The first is to develop the model through a single large effort through one RFP. A
variation on this might be a multi-stage effort in which the first RFP takes the development
process up to a certain milestone, and then another RFP issued (or even a third) in order to
complete the next stage in the project. Here the trip-based model is being used as usual while the
agency waits for the activity-based model to be ready for use.
This differs from a phased approach in which the activity-based model gradually replaces certain
trip-based model components over time. Here two model systems are not being maintained.
Rather, the trip-based model is being phased out.
A third approach is to transfer an activity-based model developed for another region and to refine
it as needed for the new location. This could be a single or multi-stage effort, though it is like to
involve multiple stages if the region is large and complex in its transportation system.
104
Page 29
Advantages
Disadvantages
New software
Entire budget must be
committed upfront
29
This slide shows the approach taken by most of the early adopters of the activity-based models.
In upfront development with a single RFP, there needs to be a large enough budget committed to
the project to pull it off in one contract. Because the model is not being transferred from
elsewhere there is control over the system design, but also some risks in developing a new mode
structure and application software. Once the work has been completed, however, the agency has
a working model that covers all of the relevant travel markets intended in the original design.
This is not to say that additional features and refinements might not be added later. In the case of
SFCTA, for example, even though they have used their CHAMP model for close to 10 years
now, they have frequently made modifications to certain model components, either for projectspecific requirements, such as New Starts or tolling analysis, or just for the sake of efficiency. In
that sense, the SFCTA modeling program is similar in its model update approach to agencies in
similar-size market areas that run trip-based models.
105
Page 30
Advantages
Disadvantages
This approach is one followed by some agencies that are committed to developing an activitybased model, but might need to stretch out the process over an extended period of time in order
to synchronize funding availability, or perhaps in order to buy time to collect new data. In these
cases, there is clearly a disadvantage to having to issue multiple RFPs, though some agencies
may be faced with little choice. Typically, these agencies will maintain their trip-based model
until the activity-based model is ready for use.
106
Page 31
Phased Development
Examples: San Diego (SANDAG), Seattle (PSRC)
Advantages
Delay some costs until
budget available
Resource development (data)
Gain familiarity with model
software and operation
Control over system design
Disadvantages
Not able to enjoy full
benefits of model design
until entire model is
implemented
31
Phased development is another option, which can help an agency spread costs and risks over
more time. At the same time, it enables modelers to get become familiar with some of the new
model components. The main disadvantage to this approach is of course delayed gratification.
The agency wont be able to fully realize the benefits of their activity-based model design until
all of the components are in place. In the case of both SANDAG and PSRC, the two agencies
began by replacing trip generation components in their trip-based model with the activity pattern
and tour generation components.
107
Page 32
Advantages
Disadvantages
If an agency wants to get started in activity based modeling, transferring a model from another
region is quick way to get started at a reasonably low cost. Listed here are examples of MPOs
that have or are in the process of developing models, based on specifications developed in other
regions. Thus far, it has seemed to work well and partially mitigated concerns over transferability
of parameters and structures. Nevertheless, an agency could be expected to follow-up the initial
transfer with model calibration and validation based on local data. Depending on agency needs
and the results of sensitivity testing, there may need to be follow-on contracts issued to refine or
redesign certain model components. This is more likely to be the case in larger, more complex
metro areas, particularly those with large-scale transit systems.
108
Page 33
109
Page 34
Budget
Development timeline
Agency
Software
Hardware
Data
Funding mechanism
34
Lets now talk about the resources required to develop and maintain activity-based models.
These include: monetary budget, the development timeline, agency staff, software, hardware,
data, and funding mechanisms.
110
Page 35
35
One of the first questions is, how much will it cost to develop the model? There are several
factors that affect the cost of a new model.
The first aspect to consider is whether you are comfortable with the paradigms that are currently
into practice. If you disagree with the foundational theory of the existing models, then part of
your development costs will include investing in the time that it takes to bring a new approach
into practice.
A second importance consideration is whether to adopt existing software, or to develop your
own? Generally speaking, agencies would prefer that the software be written in a language that
their staff is already familiar with. Current ABMs are written in various object-oriented
programming languages, such as Java or C++. None of the current models is a completely
scripted solution; that is, they do not run completely in the same commercially-available
transportation planning packages that run trip-based models. Developing new, well-tested and
debugged programs is very costly and time-consuming. Some of the software on which current
111
activity based models run is publicly available and free, and it is modular so that even if you
choose not to transfer the model itself, the software can be adapted to work with a new model.
Assuming that one chooses to transfer an existing model, a third important consideration is
whether to re-estimate all of the individual models, to re-estimate some and recalibrate the rest,
or just to recalibrate. Re-estimation is attractive because it offers the opportunity to adapt the
models more fully to the local conditions, but depending on the extent of re-estimation it can be a
large effort. A model that is transferred and only partially re-estimated can be calibrated to meet
local conditions, and this typically takes less time that re-estimating all the models.
Another aspect of transferring an existing model is whether to take as is, or to include some
additional features in the model. These additional features typically respond to some specific
need of the agency, such as refining how the model deals with road pricing, or with various
aspects of transit services. They may also be related to addressing specific populations or travel
markets that may not have been important in the original model, such as visitors, seasonal
residents, or special event travel.
Will the new model require new data? Many regions already plan to conduct travel behavior
surveys every 10 to 15 years, so a relatively recent survey may already be available. If it isnt,
then the cost of a new survey may need to be included in the cost of developing the activitybased model.
Lastly, there is the question of whether to develop the model in-house, or to hire a consultant to
do it. To date the experience in the United States has been that consultants have taken primary
responsibility for developing activity-based models. But as we will see, there have been several
instances of agency staff taking an active, hands-on role in developing parts of the model.
112
Page 36
36
Everyone wants to know how much it will cost to develop and maintain an activity-based model.
In talking to various agencies that have implemented these models, it can be difficult to pin the
answer to an exact number. While consulting costs are usually known, the budgets range widely
because models have been developed under so many different arrangements. In some places,
MPO staff members have contributed a lot to model development, particularly in developing
data. In terms of ongoing maintenance and operations, the staff members responsible for the
activity based model often have a range of other responsibilities. In addition, some the work on
land use and survey data development as well as network coding and enhancements serve
multiple purposes, including making improvements to the trip based model while the activity
based model is under development.
In addition, more recent activity-based model development projects have tended to cost a little
less than early ones. This is primarily due to the ability to transfer model components from one
region to the next.
113
Page 37
37
The SACOG model is one of the first to go into implementation. The model was paid for
primarily through a combination of grants obtained over the span of several years. The total
consulting fees are $849K, of which $514K represents the development cost of the initial model,
and $335K the cost various modeling enhancements undertaken recently. One significant
contribution of agency staff to the development of the SACOG model, not accounted for in the
consulting fee, was the development of a land use parcel database.
114
Page 38
38
San Diego is a more recent model. It has been developed in four strategic annual phases, all with
the same consultant. The consulting fee includes significant software development, as well as the
development of a series of sub-models, such as simulation of airport passengers and cross-border
travel, special events, a visitor model, and an external travel model. Total consulting fees are
$1.2M, equally distributed over 4 years. The phase 1 models were funded with a grant from
Caltrans, and partly for this reason these models were designed to result in a fully-functioning
system by the end of this phase.
115
Page 39
39
116
117
Page 40
40
Lets turn now to the time that it takes to develop a new model. Key aspects that play a role in
the development timeline are listed in this slide. They include:
Annual funding stream: can the agency secure sufficient funding upfront, or will the
model development pace need to be adjusted to fit the availability of funds?
Is there a pressing, immediate need or desire to have an activity-based model in place?
Agencies need to schedule their model update cycles to fit the cycles of regional
transportation plans and other activities. A model that is expected to be used to support an
upcoming RTP may need to be developed on a faster track.
Will new data collection be required? The time needed to design, conduct, and analyze a
household travel behavior survey should be reflected in the schedule.
The specific features desired in the final model also play a role in how long it takes to
develop the model. A model transfer without any changes to the model structure can be
performed in 6 months, but additional time is required to re-estimate certain model
118
components, add or refine model features that are important for the region, and include
simulation models of special trip markets, such as air passengers, visitors, and others.
Last but not least is the extent to which agency staff can contribute to the model
development. Some agencies may be able to take responsibility for developing certain
model components, thus effectively complementing the consulting staff.
119
Page 41
Model
Design 2001
Develop
Parcel
Database
2002-2004
Model
Estimation
2005-2006
Model
Calibration
2006-2008
Peer Review
2008
Model
Update
2011
41
This timeline was developed from information given to RSG by SACOG through a study
commissioned by the Association of Metropolitan Planning Organizations. It chronicles one of
the first activity-based models. The impetus for the development of the SACOG model came
from the MPO staff. After the initial design, there was a two-year period in which agency staff
developed an extensive parcel-based land use database. This was followed by a 3-year period of
pure model development: model estimation and calibration, as well as development of the
application software. The model has been in use since 2008. Various enhancements were
undertaken in 2011.
120
Page 42
Tour Scheduling,
Destination, Mode
Choice Models
2009
2010
Trip Models,
Calibration
2011
Model Validation,
Submodel
Development,
System Integration
2012
42
This example is a more recent model, started in 2009 and scheduled for completion this year.
The SANDAG activity-based model took advantage of substantial work that the agency had
invested in developing a micro-zone system and network GIS management tools developed for
the trip-based model. The model development was planned as a phased update. Phase 1 of this
model was funded with a grant from Caltrans, and so it was tailored to result in a fullyfunctioning model, with some AB-like components such as a population synthesizer, residential
and workplace location models, and a day pattern model, working together with the trip-based
distribution and mode choice models. There have been 3 subsequent phases, each devoted to a
specific subset of models.
121
Page 43
43
How much or how little agency staff participates in model development activities largely
depends on their interest, their skills and their availability. The end goal is for the agency to be
able to own the model and use it effectively. For this to happen, it helps to take advantage of the
model development process to get agency staff familiar with the model. Some of the desired
skills include a good understanding of the core modeling techniques, in particular discrete choice
modeling and simulation, some familiarity with the programming language on which the model
is built, and familiarity with database querying software. It helps to plan for multiple practical,
days-long training sessions as the model is developed to build familiarity with the model over
time. Comprehensive staff training is one of the lessons that agencies that have gone through this
process cite as key to success and return on their investment.
122
Page 44
Sacramento Example
Approximately 3-4 FTEs on the development of the parcel
database in 2004
4 staff working time and 3 staff working time on
modeling activities (2 FTE total)
Activity-Based Modeling: Management Institutional
44
These two examples show the extent of agency staff involvement during the development of the
San Diego and Sacramento models. Both agencies provided substantial staff time to support and
complement the consultants work.
123
Page 45
45
Like a trip-based model, activity-based models require maintenance and support activities.
Preparation of network scenarios, for example, is typically identical to the work done to run
alternatives on a trip-based model. In an activity based model there tends to be more attention
paid to land use representation at a disaggregate level. This is especially true of the models
which operate at a parcel or micro-zone level.
In addition, because these programs make use of some data structures and model forms not
typically found in commercial software, they have been developed in customized application
packages that often rely upon external back-end databases. This, coupled with the opportunity to
query disaggregate outputs in numerous ways to develop project- and policy-specific
performance measures, has made staff scripting and programming skills even more important.
That is, it is often necessary to know more than just the commercial package macro and scripting
languages to be able to fully exploit the model.
124
125
Page 46
Software
46
Activity-based models have various software requirements. They rely on the same commercial
transportation planning packages used for trip-based models for certain functions, such as
network skimming and assignment, and sometimes for running special market models. So
licenses to run your package of choice are still required. The core demand components of an
activity-based model are written in programming languages which may or may not be familiar to
an agencys staff. Some of the ABM software implementations are free; they are distributed by
the developers as open-source software. This includes the software that runs the CT-RAMP
family of activity-based models, and is expected to be the case for the work on activity-based
models that is being funded by the SHRP C-10 project. Other specialty software may also be
required to support database management, visualization tools, and distributed computing
solutions.
When it comes to software, the issue for agencies is not so much the licensing costs, which in
many cases are zero, but the need to have staff fluent in these types of software tools. Other
software issues to think about are related to providing remote access to the agencys servers to
126
third party users (partner agencies, consultants working on behalf of the agency). Remote access
and/or cloud computing solutions are more critical for large regions, where multiple, powerful
servers are required to complete a model run in less than a day.
127
Page 47
47
In a trip-based model run time is approximately proportional to the square of the number of
zones; add zones and the run time grows exponentially. In an activity based model, run time is
proportional to the size of the population; fortunately when you add population the run time
increases linearly. The key message though is that populous regions incur longer run times, all
else equal, than smaller regions.
How much to spend on computers to run the model is directly related to how fast youd like the
model to run. Most, if not all, models can be configured to run on a single, multi-processor
computer. Most everyone would like to see their model complete a run in less than 12 hrs or
approximately overnight; start the model when you leave for the day and have the results ready
for you when you come back the next day. The way to achieve these run times in a large region
is to deploy as many processors as needed to achieve the desired run times, either on a single
computer or distributed over multiple machines. And as you all know, the more powerful the
computer needed, or the more computers needed per model run, the higher the cost of the
hardware. The good news is that computers continue to get faster and faster, and less and less
128
expensive over time. Also ABM developers continue to come up with strategies to optimize the
software, sometimes achieving significant improvements in run time performance.
Page 48
48
This slide shows two comparisons of run time performance between trip-based and activitybased models. SANDAG achieves approximately the same performance with their activity-based
model as they do with their trip-based model. It should be noted that the SANDAG trip-based
model operates on the same micro-zone transit access framework that the activity-based model
does. In Sacramento, the activity-based model takes approximately 4 times as long as the tripbased model, but it operates at a parcel level while the trip-based model operates at a far more
aggregate traffic-analysis zone level. In both cases, the activity-based model provides a far more
detailed representation of travel demand than the trip-based model, so in many ways the models
are not really comparable.
129
Page 49
49
An activity-based modeling system does not necessarily have to be slower than a trip-based
modeling system. In this particular case of a new DaySim model being developed for Fresno,
California, the activity-based modeling system runs faster than the trip-based model, on a
machine thats actually slightly slower (fewer cores, slower RAM)
130
Page 50
Data Requirements
Data requirements are the same or similar to those of
trip-based models
Some optional model features call for additional data
collection:
Parcel or micro-zone population and land use inventories
Parking availability, transponder ownership, transit pass
ownership
Highway and transit operations data for multiple time periods
50
Another often-voiced concerned about activity-based models is that they are data hogs. In fact,
an activity-based model can be developed with the exact same data that are used to develop tripbased models: a household travel survey, transit on-board survey, traffic counts, transit
boardings, census summary data, employment data, and origin-destination or intercept surveys
when available, for example.
Some optional model features call for additional data collection. Models that operate at the
parcel or micro-zone level require detailed land use inventories. Models that account for mobility
attributes like parking availability and cost at the place residence or place of work, transponder
ownership, or transit pass ownership require data on who has access to these facilities and it
impacts their travel choices. Agencies that wish to model highway or transit operations in great
temporal detail need to be prepared to maintain a larger set of network attributes in their
databases, for example.
131
132
Page 51
Data Requirements
Recent household survey required for model estimation
and development of some calibration targets
Activity based modeling is less forgiving of incomplete
person roster, trip diaries or missing information
Requires consistency across trip choice dimensions and across
individuals
But it can make use of data that is typically asked for but not
used by trip-based models
Age, gender, occupation, employment status, driver license, usual
workplace and school locations, vehicle used, etc.
51
Household travel behavior surveys are the backbone of all model development, and as mentioned
previously, activity-based models use the same household surveys that trip-based models do.
Nonetheless, it must be mentioned that an activity-based model is less forgiving of incomplete or
inaccurate information. The main issue is that the model requires completeness and consistency
across all individuals in the household and across the trip choice dimensions. For example, if
joint travel among household members is a feature of the model, then the survey used to estimate
the joint travel components must be able to show which household members travel together.
Often times people report that they traveled with other household members, but their reported
trip schedules or destinations do not coincide.
While activity-based models are more demanding of the data, they also use more of the data that
is reported in these surveys. Person attributes are rarely used in trip-based models, yet are
common and powerful explanatory variables in activity-based models. Activity-based models
care about long term choices such as usual work place and school locations, in addition to short
term choices did you travel to work or school on the survey day? The ABM can track which
133
person in the household uses which vehicles, which potentially can be used for analysis of fuel
consumption and adoption of new vehicle technologies.
Page 52
Funding Approaches
Build into model development work program
External grants (SACOG, SANDAG)
In-kind, cost-sharing arrangements
MPO staff develop land use database, networks, auxiliary
demand (SANDAG)
MPO staff develop enterprise database, software (DRCOG)
52
Last but not least, where to find the money to pay for all this? The most common approach is to
funding it via the agencys ordinary work plan, at least by partially diverting some funds that
would have otherwise been invested in improving the trip-based model. Thus, one opportunity to
introduce an activity-based model is when your trip-based model is due for a major overhaul.
A second approach has been to fund it via external grants. SACOG and SANDAG were both
able to secure grants from Caltrans to fund part of their activity-based model program.
A third option has been to fund the development partially with in-kind services performed by
agency staff. In this way the funds go to pay for your staffs time, ather than for consulting fees.
A fourth option that has become feasible now that multiple agencies share a common modeling
framework is to jointly fund part of the model development.
134
The importance of a local champion and early success for securing continuing funding cannot be
overstated. In this respect, it may be more important for the long term success of the program to
start small and demonstrate the usefulness of the model with some early applications, than to
spend years and years developing the most advanced, custom-built tool possible.
Page 53
53
In the 10 year or so that activity-based models have been in practice, what has been learned in
terms of user experience? From the perspective of the user, how are these models different from
trip-based models? We are going to talk about four aspects of this user experience:
Page 54
54
increase confidence among the staff and stakeholders that the model provides reasonable and
relevant answers.
137
Page 55
Model Applications
SFCTA Applications
55
You would think that with nearly 10 years of activity-based models being implemented and in
application that there would be a long list of projects that have been performed with these
models. If you think that, then you are absolutely right. This slide shows a sample of the projects
and studies that have been performed with the SFCTA ABM, which you may recall was the first
operational model in the United States. Since 2001 SFCTA has used their model to develop their
congestion management program and countywide transportation plan, to perform multiple transit
studies include BRT on key arterial roadways, the Third Street light rail rail study and more
recently the analysis of alternatives for the Central Subway, which supported their FTA New
Starts Application. One of the pioneering studies performed with the SFCTA model was the San
Francisco Mobility Access and Pricing Study, which examined alternatives for charging for auto
access into the San Francisco central business district.
138
Page 56
Model Applications
NYMTC Example
56
The New York Best Practice Model is another example of an early model -- in operation since
2002. The BPM has been used extensively to support multiple projects and studies for several
agencies in New York City, as well as air quality conformity and the regional transportation
plan. This slides shows a small sample of BPM applications. One of the most highly visible
transportation studies in the country, the Manhattan Area Pricing Study, was performed with the
NYBPM. But it has also supported the types of projects that may be more common in other large
metropolitan areas, including various toll studies, multiple subarea studies that examined local
road and transit needs, and various transit projects, including the Tappan Zee Bridge New Starts
alternatives analysis.
139
Page 57
Model Applications
SACOG Example
2 Air Quality Conformity Reports since 2008
2010 SB375 greenhouse gas (GHG) emissions analysis
2008 head-to-head comparison with SACMET (trip-based
model) in developing the 2035 Metropolitan Transportation
Plan
Placer Vineyards transit-oriented development scenario
analysis
Curtis Park Village infill development project scenario
analysis
57
The SACOG activity-based model has been in operation since 2008 and already has multiple
projects and interesting applications under its belt. The model has been used to prepare two air
quality conformity reports since 2008, and the emission analysis called for by Californias SB375 regulation. One of motivators for developing an activity-based model in Sacramento was to
study alternative land uses and formulate a long term land use strategy. Two of the applications
listed here, Placer Vineyards TOD scenario analysis and the Curtis Village infill development
scenario analysis exemplify these types of land use analyses.
140
Page 58
Model Applications
Oregon Statewide Model
Oregon Bridge Study
Oregon Statewide Freight Plan
Willamette Valley Land Use and Transportation Visioning
Study
58
The two statewide activity-based models have been largely absent from our discussion thus far,
but they too have been put to good use. SWIM, the Oregon Statewide model, or TRANUS, its
precursor, have been instrumental in supporting various statewide studies. The Oregon Freight
Plan, for example, examined the impact of various economic scenarios on freight movements
across the state. The Ohio Statewide model also has been used to study various projects of
regional significance, such as the economic impact of the proposed US 22/36 highway on the
eastern side of the state, and a study of commodity flows over the Brent Spence bridge, which
spans the Ohio River at Cincinnati.
141
Page 59
59
Travel demand models have multiple constituencies. The agency that developed and owns the
model often times makes it available to other municipalities, transit agencies, and consultants
acting on behalf of these institutions. Making sure that all these external users are comfortable
with and able to use the model is a key step in the process of ensuring that the model is useful
and relevant for a variety of stakeholders. As is often the case with new technologies, there may
be resistance to adopting the model for a variety of reasons, including lack of familiarity,
skepticism about claims that it is in fact a better tool, and concerns about the cost and time
required to bring staff up to speed. The keys to success are hands-on training, extensive
documentation that covers the model fundamentals as well as its operations, and user-friendly
ways to interface with the model, both while preparing input data and scenarios and when
analyzing its outputs.
142
Page 60
SANDAG
Provides remote access to its servers
ARC
Cloud computing implementation for external users
60
There are success stories that suggest that none of the concerns that we just discussed are
insurmountable barriers. The long list of applications that I just showed you is proof that these
models can be used by agency staff and others in the course of their everyday work. In the 10
years that it has been in practice, the New York Best Practice Model counts more than 30
different users, including consultants and local agencies other than NYMTC.
One of the obstacles cited towards acceptance of some activity-based models is the need to own
a cluster of computers in order to run the model in a reasonable amount of time, which can be a
substantial cost for infrequent users of the model. Agencies such as ARC in Atlanta and
SANDAG are exploring ways to provide easy access to their model, whether via remote access
to their own servers, or by making the models available in the cloud.
143
Page 61
61
Lets turn now to the last topic of todays presentation, stakeholder acceptance and use. The key
to acceptance of the model by stakeholders, both inside and outside the agency, is to focus on
answering the questions that are relevant to the agency. Because the models produce output that
is akin to data from a travel survey a list of people with information about all their trips they
provide unprecedented opportunities for data exploration and derivation of performance
measures. The main concepts behind the models are easier to grasp by lay people because they
relate closer to their own behavior than trip-based constructs do. But to the unprepared staff
person, the models can appear as double-edged swords, in that the vast amount of model output
can be challenging to sort through and summarize in clear, concise ways that tell a story. So
again, one key to model acceptance is to make sure that your staff understands it well, so they
can explain it to others and use it effectively. A second lesson learned is that it pays to develop
tools and procedures to prepare standardize reports to present results in visually appealing ways.
144
The following slides show various examples of reports and charts produced with the Atlanta
visualization tool. This tool was originally developed for the Oregon Statewide Model, and since
then has been enhanced and applied as part of other model systems.
Page 62
62
The Atlanta ABMViz is a visualization dashboard. Its a stand-alone tool that includes multiple
built-in standard reports, as well as the ability to build queries of the datasets produced by an
activity-based model. It can be used to compare performance measures across scenarios or across
regional subareas. It generates a variety of built-in reports, from simple one-way and two-way
tables to charts, maps and animations.
145
Page 63
63
This slide shows how time is used over the course of a day. The drop down bar at the top lets you
select different person types full time workers, or university students, of pre-school children,
for example. The colors identify different types of activities at home, at work, at school, etc.
The chart shows that, at noon for example, about 40% of the population is at home, 13% of the
people are at school, and over 25% are at work. This type of chart can be useful, for example, to
show the effect of telecommuting policies on work at home and on the time use of other family
members. Some effects of telecommuting incentives on workers are obvious theyll be more
likely to stay at home and work from there. But others are less so will the gain of time that
was formerly used for traveling now result in more time spent working, or more time spent in
non-work out of home activities? What will be the impact of people working from home on the
time use of other family members, for example?
146
Page 64
64
This type of chart is called a radar chart. It is used to compare multiple performance measures
across different population groups, which can be identified by geographic location (as shown in
this example), or by a population attribute (such as household income, or auto ownership, or
ethnic group). The four corners of each blue diamond correspond to four different performance
measures. In this example the performance measures are jobs-housing balance, transit mode
share, accessibility, and zero car transit trips per household. The orange area shows how well
each population group does with respect to each of the four measures. Where the group does
well, the orange area touches the corresponding corner. The size of the orange area is relative to
how well the group scores relative to all other groups being compared. ARC has found that these
types of charts are useful in planning studies to show and explain the impact of various scenarios
on different population groups.
147
Page 65
Ongoing Developments
65
Well end our discussion today with a brief overview of whats going on with the
implementation of activity based models.
First of all, now that there are two relatively mature and well-tested model systems, DaySim and
CT-RAMP, we are seeing multiple instances of model transfers, typically coupled with some
adaptations. There is continuous improvement of existing designs, whether to incorporate
research findings or to address populations and travel markets that were somewhat ignored or not
well-understood previously. There are new paradigms being put into practice, as is the case in
Portland and Los Angeles. Hardware and software continue to improve, resulting in better model
performance. There is a lot of interest and on-going activity towards integrating activity-based
models with dynamic traffic assignment, as well as towards integrating urban land use models
with activity-based models.
148
Page 66
66
149
Page 67
150
Page 68
February 2
February 23
March 15
April 5
April 26
May 17
June 7
June 28
July 19
August 9
August 30
September 20
68
151
Page 69
69
152
153
daily travel. If we have a policy for transit support for elderly people, we know which individuals
are over the age of 65 and how their travel changes.
Rosella: There are many policies that we use trip-based models for because we have to, but there
is some dissatisfaction with the results because we know the tool is overly sensitive or not quite
right to measure the question.
Do travel surveys at statewide level make an effort to oversample for inter-city trips or visitor
trips? How are long-distance trips captured?
Rosella: For inter-city travel in Ohio, they had three different urban area surveys by the time they
did activity based model. They also did a statewide study to cover other parts of the state. They
followed up with some respondents to get information about long distance trips within a certain
time period, maybe 15 days.
If activity based models are going to be used to measure effects from higher gas prices and
decreased household budgets, what basis is used to determine household behavioral responses?
John: That's a cutting edge model application. The response is captured in a household's
propensity to under-generate tours. If costs go up, they travel less. What isn't so explicit is what
they are trading off when they spend more money on gas.
Rosella: As part of some sensitivity tests that Oregon did with their statewide model, they looked
at effects of increased gas prices. There was a paper presented at TRB about the responses to
price increases.
Does agent based modeling differ from activity based modeling? If so, how?
John: Agent-based modeling actually has a very particular meaning. There is a field called gentbased modeling in which people look at emerging complex systems. They will look at agents,
who are entities who respond to simple sets of rules and interact with their neighbors. An
example is SWARM modeling. Another example is traffic flow modeling. In this sense, this is
very different than activity based modeling where we are micro-simulating individuals with a
prescribed aggregate behavior. We dont wait for complex patterns to emerge. In fact, we model
the patterns directly and calibrate the models to behave a certain way.
What are typical travel survey costs, by rate per completed survey?
John: At this time the average is about $150 per completed household survey. If GPS units are
involved, it could be more.
Specific to slide on San Diego model cost, did that value include the survey?
Rosella: No.
154
For an agency staff with limited time and resources, what areas of model development should
staff attempt to get more involved in?
Rosella: What typically happens is that the agency staff takes responsibility for procuring input
data. They also may take care of network and land use data. That is very common and is a good
way to share work since they are often more familiar with the particular geographic area than the
consulting staff may be. Beyond that, it depends on the skills and interests of staff. In San Diego,
staff developed the Population Synthesizer. In Denver, agency staff members wrote the entire
software package. There is a wide variety of involvement out there, and those are two very large
agencies.
From someone who works at RTD in Denver, in the context of doing model application inhouse. Can he get contact information from other places that have implemented an activity based
model? Can he also get information on how the MPOs and Transit agencies have shared
modeling resources?
Rosella: Not all models have been created by MPOs or transit agencies. I am not entirely familiar
with all the users of the New York model, but can find more information after webinar.
John: TMIP has online discussion forum and that is a good resource for making contacts.
Maren: I wouldn't see a problem with sharing contact information for all the agencies included in
presentation, but will find out.
What is the typical run time for a model running every year or every five years?
Rosella: Run time that was reported in presentation was the time for one scenario, with all
feedback, from beginning to end. It's not related to how many scenarios you want to analyze and
for how many years. That's a different topic.
John: It's also different for models that have a land use component.
Can we use activity based models for data from other regions? Are there MPOs who would share
their model for study?
Rosella: In terms of sharing models, probably better addressed by the owners of the model. In
terms of sharing data, University of Minnesota has a repository that goes back several years. If
you're looking for data to explore, that's a good option. For a specific region, you should contact
model owners.
155
156
Page 1
Activity-Based Modeling
Session 3: Technical Issues for Managers
157
Page 2
Acknowledgments
Media Production
Bhargava Sana
Brian Grady
Activity-Based Modeling: Technical Issues For Managers
Resource Systems Group and Parsons Brinckerhoff have developed these webinars
collaboratively, and we will be presenting each webinar together. Here is a list of the persons
involved in producing todays session.
Joel Freedman and Maren Outwater are co-presenters. They were also primarily
responsible for preparing the material presented in this session.
Stephen Lawe is the session moderator.
Content development was also provided by John Gliebe.
Bhargava Sana and Brian Grady were responsible for media production, including setting
up and managing the webinar presentation.
158
Page 3
February 2
February 23
March 15
April 5
April 26
May 17
June 7
June 28
July 19
August 9
August 30
September 20
3
For your reference, here is a list of all of the webinars topics and dates that have been planned.
As you can see, we will be presenting a different webinar every three weeks. Three weeks ago,
we covered the second in topic in the seriesInstitutional Topics for Managers. Today we will
cover technical issues for managers. Todays session provides a high-level overview of technical
issues for management staff contemplating the development or implementation of activity-based
models for their region. In three weeks, we will begin to cover the details of activity-based
models, including choice models, simulation, and different activity-based modeling frameworks.
159
Page 4
First, wed like to summarize the topics covered in the first two sessions in the series. Session 1,
the Executive Perspective, was intended to provide a high-level overview of activity-based
modeling and how it could benefit an agency involved in policy studies. You may find a
recording of Session 1 in the archives. In this session, we began by discussing what an activitybased model is and contrasting it with a trip-based model. We outlined some of the practical
advantages and limitations of activity-based models. After that, we described a several policy
examples in which activity-based models have been used to analyze some fairly complex policy
scenarios, including congestion pricing, travel demand management, land use interactions, and
latent demand. We concluded this first session by touching briefly on staffing, consultant, and
hardware and software resource issues, and shared some lessons learned from our experience in
model development projects.
160
Page 5
The second webinar series covered institutional topics for managers. In this session, we began be
describing the advent of travel modeling in the U.S., how the needs and resources of modelers in
the early days dictated trip-based model design. We then describe how models began to evolve in
sophistication, sometimes in response to challenges to their authenticity. We then began to
describe the advantages that activity-based models provide in terms of information content and
how that information can be used in policy analysis. With these advantages in mind, we then
outlined the questions typically on the minds of persons who are contemplated the move to an
activity-based modeling systemcosts, performance, resource requirements, and data. We then
listed 17 locations in the U.S. where activity-based models have been developed or are under
development, and described three different approaches to model developmentupfront, phased,
and transferred.
The second half of the second webinar provided several real-life examples of agencies
development costs, timelines, hardware and data requirements, and funding approaches. We then
161
concluded with a detailed discussion of user experiences, including calibration and validation,
examples of model applications, stakeholder acceptance and usage, and communicating results.
162
Page 6
In todays session, we would like to accomplish two broad objectives. First, we would like to
cover issues that we feel are important to managing the technical processes involved in activitybased model development and usage. By the end of this session, you should be able to discuss
the processes used to develop, calibrate, validate and implement an activity-based modeling
system.
The second learning objective is to discuss the criteria that agencies should consider when
evaluating whether an activity-based modeling system is something they should pursue. This will
be our attempt to tie a bow around the discussion of management issues related to activitybased modeling.
The third learning objective is to discuss the high-level model design decisions that will need to
be made when embarking on activity-based model project. If an agency decides to postpone
development of an activity-based model, what are some of the alternative transitional
163
development paths that an agency can take in enhancing its modeling capabilities? It is our intent
to answer this question, too.
164
Page 7
Terminology
Model development
Model estimation
Model calibration
Model validation
Model implementation
Model application
In previous sessions, weve defined some of the terms used to describe what an activity-based
model is. In future technical sessions, we discuss much more of those details. For this session,
however, wed like to define some terms related to the processes of model development and use.
Model development and analysis consists of three analytical activities: estimation, calibration
and validation. Each of these activities may be characterized by different sets of methods aimed
at fitting model outputs to observed data.
Estimation refers to the process of developing coefficient estimates for explanatory variables in
individual model components, usually based on samples of individuals or households, such as a
found in a household survey.
Calibration refers to applying the model to a set of inputs, usually representing the entire
modeled population; comparing outputs to key benchmark values, derived from expanded survey
data; and adjusting certain model components, iteratively, until a desired level of fit is attained.
In some cases, it may be desirable to calibrate models to independent sources of data. A common
165
example of this is to calibrate the means distances of workplace location choice models to
Census journey-to-work data.
Validation refers to comparison of model outputs to independent observations, such as traffic
counts, transit boardings, and Census-derived characteristics of the population.
Model implementation is a term that is used in different contexts. For some, model
implementation refers to the transformation of a mathematical model into an application program
that can be run. In this sense, implementation is a preliminary step that allows one to begin
model calibration. In another sense of the term, model implementation refers to the integration of
an activity-based model into a larger forecasting system, which includes all demand model
components, network models and auxiliary demand, prior to model validation. These are the two
ways in which we discuss model implementation in this webinar.
In yet another sense, agencies may say that their model has been implemented, indicating that
they have reached the milestone event of being able to use that model in a production context,
which often implies that it has also been validated.
Model application is used a little less ambiguously. Typically, this refers to running the model
to produce some type of analysis.
166
Page 8
Synthetic
Population
Auto
ownership
Transit pass
Mobility
Choices
Trips
Destinations
Modes
Model
Inputs
Long-Term
Choices
Daily
Activity
Patterns
Trip
Assignment
Model
Outputs
Activity-Based Modeling: Technical Issues For Managers
Work location
School
location
Tours
Purposes
Schedule
Highway
volumes
Transit
boardings
For purposes of discussion in the remainder of this webinar, it might be useful to review the
basic components of an activity-based modeling system. While activity-based models can vary in
structure, this diagram shows the location of tour and trip detail choices (tour mode, primary
destination, intermediate stop location and trip mode) in a typical model stream. The text on the
outside shows the types of outcomes predicted by each model stage. When we are ready to
predict tour and trip details, we already have a synthetic population of households and persons
with their home locations; we have predicted the primary work and school locations, auto
ownership and other mobility decisions; and we have generated and scheduled tours using a daily
activity pattern model. We do not yet know the primary destination of any non-work and nonschool tours, the tour mode, the location of intermediate stops, or the trip mode. Once we are
able to fill in these details, we are ready to convert the simulation data into trip tables that can be
assigned to the network.
167
168
Page 9
Model
Estimation
Activity-Based
Models:
More estimation, less
calibration
Model
Implementation
Model
Calibration
Model
Validation
Trip-based models are typically developed in a linear fashion. First models are estimated using
statistical techniques. For example, trip rates are calculated for household segments. Then
models are implemented in application software. Then each model is calibrated, or adjusted, to
better match expanded survey data. For example, friction factors are adjusted in gravity models
to better match trip length frequency distributions. Mode-specific constants are adjusted to better
match trips by mode. Finally, model outputs, such as link volumes and/or transit boardings, are
compared to independent data to judge the validity of the system.
The process of developing activity-based models is very similar. However, because choices
models are the foundation of most activity-based model systems, there is a greater reliance on
model estimation rather than aggregate calibration. Also, since models are more inter-connected,
often estimation begins with the last model in the sequence, such as trip mode choice. Then the
model is implemented and the information from that model is used to estimate the next higher
level in the model chain. Activity-based models can be validated in traditional ways; however, it
169
is also possible to run the data that was used to estimate the models through the system to check
the model predictions against the survey outcomes.
170
Page 10
Activity-Based
Model System
Structure
Pre-Processors
Network Preparation, Data Formatting, etc.
(Atlanta Regional
Commission)
Other Demand
EE Demand, Commercial Vehicle Model
Skims
Feedback Loops
Activity-based demand
component (CT-RAMP)
Transit Assignments
[AM, PM, off peak] * [Walk, Drive] * [Local,
Premium]
Post Processors
Performance Measures, Value Pricing, Emissions, etc.
10
This diagram is a typical example of where an activity-based demand model would be integrated
into a larger model system run. This is an example from the Atlanta Regional Commission. It
might be useful for you to keep this diagram in mind as we discuss model calibration and
validation.
As you can see, many of the steps you see here are the same or quite similar to what you would
have in a trip-based modeling system. This includes the steps typically found in a model run
stream, including: pre-processing network and other data, network and skim building, bringing in
static auxiliary demand tables from other sources, and then entering the central model loop.
Within the feedback, the activity-based demand model is run, along with any other models that
might be sensitive to congestion effects. The feedback loop would include intermediate highway
(and sometimes) transit assignments and would be run iteratively until convergence. Once the
feedback loop has converged, final transit and highway assignments are run, and any postprocessing modules are applied.
171
172
Page 11
On-Board survey
Difficult to use for ABM because only trip data is reported
Special questions required to construct tours
Origin-destination
Downtown parking
Visitors
Special events
Air passengers
Other special markets/populations
Census data
Calibration summaries
Synthetic population controls
Activity-Based Modeling: Technical Issues For Managers
11
A household travel survey provides the information required to estimate an activity-based model.
Several other types of surveys provide information useful for estimating and calibrating the
model.
Household survey: The household survey ideally records all travel by all household members
that occurs on a designated travel day. Required information for each trip includes location, party
composition, start and end times, purpose and mode. In addition to the trip information, the
survey compiles information about the household and its members, such as income, number of
vehicles owned, usual workplace and school locations, and availability of free or subsidized
parking at the workplace. Care should be taken to perform, in real-time, adequate consistency
checks on the information provided, to ensure high quality data for constructing complete tours,
joint tours, and activity scheduling and geo-coding.
On-Board survey: On-board survey data is most often used in activity-based model calibration.
On-board surveys are difficult to use for estimating activity-based models because typically only
173
one trip out of the entire tour is reported. Questions can be added to the survey form to elicit
information that allows inferring the entire tour. An alternative is to over-sample transit users
when conducting the household survey (though this likely provides only one transit user in the
household).
Special market surveys: Various targeted surveys can be conducted to understand and gather
information about special markets of significance in the model region, such as visitors, air
passengers, downtown parking users, and others. The importance of these markets varies from
region to region and may be relevant only for particular transportation projects.
Census data: The Decennial Census and American Community Survey provide information
critical for developing the synthetic population controls and for calibrating the auto ownership
and workplace location models.
174
Page 12
Housing units
Employment and square footage by industry group
School enrollment by grade level, college/university
Urban form and open space
12
Land use data: Land use data characterize the location and type of activities in the model
region. They may be gathered from many different sources, including census, general plans, real
estate appraisers databases, departments of education, payroll tax databases, and other sources
of employment data. Similar data is used in trip-based models. Activity-based models that have
been designed to work on parcel or micro-zones require that the land use data be expressed at
these levels of geographic aggregation, instead of the larger traffic analysis zones.
Parking supply data: Like trip-based models, activity-based models require an inventory of
parking lot locations and prices for parking-constrained areas.
175
Page 13
Estimation Files
Vehicles
Activities
Trip Level
Models
Trips
Tour Level
Models
Tours
Persons
Households
Individual Daily
Patterns &
Schedules
Person Day
Level
Models
Joint
Tours
Household
Daily Patterns
& Schedules
Joint Tour
Level
Models
LU
LOS
Household
Day Level
Models
Long-term
models
13
Household survey data is the foundation of activity-based models. Survey data typically consists
of household, person, and trip (or activity) files, plus a vehicle file. These survey data files, are
shown on the left. They are assembled into core data files, as shown in the middle of the graphic.
They include tour files and daily activity pattern files that describe what each person did on the
surveyed day. These are the observed data files that are used to calibrate the system. A more
comprehensive set of data files, as shown on the right, is required for model estimation. This
involves appending land-use and level-of-service data to the core data files, typically with one
record per choice with descriptions of each alternative that was available for that choice. These
files are read by estimation software.
176
Page 14
14
Model estimation is performed in statistical software (SAS, SPSS, Stata, R, Alogit, NLogit,
Biogeme) by an experienced analyst. Typically models are estimated in order from bottom up;
that is, the last model to be applied is estimated first. After each model is estimated, it is
implemented so that the expected utilities (accessibilities) from the model can be used in upperlevel models. Once all the models have been implemented, the entire system can be calibrated.
177
Page 15
Tour time-of-day
Entire-tour mode
Stop frequency
Stop location
Accessibility of potential
activity sites along the
route between the primary
destination and home
Trip mode
Trip departure time
Activity-Based Modeling: Technical Issues For Managers
15
This graphs illustrates a portion of an activity-based model system in which tour-level decisions
condition trip level choices. The model is applied from the top-down, with tour-level choices
made first. In order to estimate the upstream models, however, it is necessary to first estimate the
models downstream. This is particularly important when the models are structured as a series of
nested choices, which most activity-based models do. The composite utilities (log-sums) of
lower level choices are used as explanatory variables in the choices above them.
This might be the ideal process for model development. Lets take a look at some other options.
178
Page 16
16
In the second webinar we talked about alternative paths to developing activity based modeling
systems. One of these options was to transfer a model system from another region. That is, we
would borrow model structure and initial parameter estimates. For certain model components,
such as activity generation modules, there tends to be a high degree of transferability, because
household and person types have the most influence on model parameters and people tend to be
similar across regions. Other model components, however, such as tour or trip destination
choices, are heavily driven by the unique spatial structure and transportation system service
characteristics of each region, and really require estimation on locally derived data. This might
be a refinement step that follows the initial model transfer. For all model components, calibration
based on available local data is highly recommended.
179
Page 17
Person 1
Attributes
Age
Gender
Employment Status
Student Status
Work Location
School Location
Tour 1
Person 2
Tour 2
Choice Models
Tour 3
Synthetic Population
Person 3
Tour 4
Tour 5
Attributes
Number of
Stops
Tour Mode
Persons
on Tour
Primary
Destination
17
Before we can calibrate a model, we need to code it into an application program. As mentioned
in the first webinar, activity-based models utilize some form of simulation to predict outcomes.
Remember, we are simulating the movements and decisions of persons and households. The tripbased modeling structure of creating numerous matrix files, simply does not work in an activitybased modeling system in which you have to track individuals.
Most activity-based models rely upon object-oriented software for implementation. A paradigm
referred to as composition in object oriented software are used to represent relationships between
data items, such as households, which have attributes such as household size, autos owned, and
possibly other attributes. Persons belong to households, and are also attributed with certain
characteristics. Persons make tours, which consist of trips. Note that some of these attributes are
read from the synthetic population, and some of these attributes are filled in by choice models.
180
Page 18
Household
Data
Manager
Thread 1: households
1-50,000
Thread 2: households
50,001-100,000
Thread 3: households
100,001-150,000
Thread 4: households
150,001-200,000
18
The ability to process each household separately makes it easy to run activity-based models
across multiple processors. Often, models are threaded to take advantage of multiple cores on a
single computer. In this example, each of four cores are tasked with performing workplace
location choice for a set of 50,000 households.
181
Page 19
MasterNode
TaskScheduler
Results
ResultsWriter
19
If even more computing power is needed (or desired), models can also be distributed across
multiple computers to decrease run times (or allow for greater complexity, disaggregation, or
behavioral fidelity in the model system). This graphic shows how a single computer sends
bundles of households to a cluster of worker machines, which then process the choice models on
those households and returns the results to the master computer which writes those results to
disk.
182
Page 20
Model Calibration
Once a model component is implemented, the synthetic
population can be run through the software, and the model
predictions are compared to expanded data
Alternatively, household survey data can be run through the
software instead of the synthetic population
Models can be calibrated in reverse order
Upstream model errors can be eliminated from the calibration process
20
Model components are typically implemented in stand-alone software such as C++, C#, or Java.
Once a model component is implemented, the synthetic population can be run through the
software, and the model predictions are compared to expanded data. Model parameters
(alternative-specific constants) can be adjusted so that predictions to better match observed data.
Alternatively, household survey data can be run through the software instead of the synthetic
population. This provides the ability to calibrate models from bottom-up, rather than top-down,
because the survey has the observed choices for all of the upstream model components. In this
way, upstream model errors can be eliminated from the calibration process. Model calibration
should focus on meaningful, defensible adjustments to model parameters.
183
Page 21
Model Calibration
Similar calibration comparisons as in a trip-based model
Trip generation rates -> tours and stops by purpose
Trip distribution -> tour destination choice and intermediate stop
location choice
Mode choice -> tour and trip mode choice
HB work trips -> usual workplace (Census JTW data)
21
The calibration of a activity-based model is quite similar to that of a trip-based model. The goal
of the calibration is to verify that the model reproduces certain patterns of travel behavior, and
there are parallels between the types of summaries used to calibrate a trip-based model and the
summaries used to calibrate an activity-based model. There are, however, some key differences:
184
Page 22
22
Activity-based models rely heavily upon person-level characteristics, such as work status, age,
and household role, in explanatory variables. Often, activity-based models gauge model
goodness-of-fit based upon these market segments. For example, does the model send the right
number of full-time workers to work on an average weekday?
It is very important that the synthetic population and the calibration target values derived from
expanded survey data agree in terms of attributes and attribute levels (e.g., age, income groups,
auto sufficiency, etc.) before models are calibrated. Otherwise, the model calibration summaries
will be difficult to interpret and compare to the household survey. This slide shows a summary
that was put together for the Seattle activity-based model, comparing the number of persons by
person-type in the synthetic population to the household survey. It shows a very good fit between
the two data sets by person type. Additional summaries would include household attribute
distributions (such as households by income, size, etc) and geographic distributions (such as
households by district).
185
186
Page 23
Mandatory
Non Mandatory
Home
100.0%
90.0%
80.0%
70.0%
60.0%
50.0%
workbased
soc/rec
meal
40.0%
shop
30.0%
pers.bus
20.0%
escort
10.0%
0.0%
school
work
23
Model calibration of tour generation model components includes both the overall daily activity
pattern and the numbers of tours by purpose. These calibration checks are often performed by
person type. The chart at the top left of the screen shows a calibration summary for a daily
activity pattern model whose alternatives are mandatory (at least one out-of-home mandatory
tour), non-mandatory (at least one out-of-home mandatory tour) and stay-at-home. The summary
shows the share of persons by type who engage in each of these activity patterns. A similar
comparison would be created for the observed data, and constant terms can be adjusted to
improve fit.
187
Page 24
Calibration of Tours
40.0%
35.0%
Observed
30.0%
Estimated
25.0%
20.0%
15.0%
10.0%
5.0%
0.0%
escort pers.bus
shop
meal
soc/rec
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Model
Survey
1 Work 2 Work
1
2
Wrk &
School School Schl
24
Once the daily activity pattern model is calibrated, the exact number of mandatory tours can be
calibrated by person-type, in order to ensure that the right numbers of tours are generated.
188
Page 25
Calibrated
0.0000
-0.8884
-0.9572
-0.8739
-1.0212
-0.3650
-0.4576
-1.1964
-0.3567
-1.2451
-1.3139
-1.2306
-1.3778
-0.7217
-0.8143
-1.5531
1.0705
0.0000
-1.3157
1.1701
-0.3000
-1.1157
-2.8699
0.0000
2.2136
2.6766
2.5194
0.0000
-3.3350
-1.8000
0.2136
4.6766
3.5194
0.3500
Escort Tour
Alternative Specific Constant
Full-time worker
Part-time worker
Personal Business Tour
Alternative Specific Constant
Full-time worker
Part-time worker
Shop Tour
Alternative Specific Constant
Part-time worker
Meal Tour
Alternative Specific Constant
Full-time worker
Part-time worker
Social / Recreational Tours
Alternative Specific Constant
Full-time worker
Part-time worker
Interaction Effects
Work + Shop
Work + Meal
Personal Business + Shop
Estimated
Calibrated
-3.5706
0.0000
0.2330
-3.5321
-0.0400
0.3830
-1.9002
0.0000
0.1670
-2.0880
-0.0400
0.3170
-3.1667
0.3400
-2.9325
0.4900
-4.9213
0.0000
0.0000
-3.8873
-0.0400
0.2000
-2.3216
0.0000
0.1241
-2.3718
-0.0400
0.2741
-1.0889
-0.6221
-0.2805
-0.5889
-0.1221
-0.0805
25
This slide is an example of alternative-specific constant adjustment for the DaySim daily activity
pattern model developed for Seattle. There are a number of choice dimensions handled by this
model, including number of tours by purpose, and number of stops by purpose. This table
summarizes the adjustments made to the alternative-specific constants that represent the number
of tours by purpose. There are constants for the core alternatives, such as work, school, escort
tours, and other tour purposes, as well as constants by person-type and tour purpose.
189
Page 26
200,000
150,000
estimated
linear trend
100,000
50,000
50,000 Normalized
100,000CTPP Worker
150,000Flows200,000
250,000
26
Destination choice models need to be calibrated for both primary tour destination as well as
intermediate stop locations. Shown on this slide are some model calibration summaries for the
San Diego work location choice model. The results were compared to census worker flows at a
district level (as shown on the left) as well as trip length frequency distributions (as shown on the
right). Note that the model did not require any K-factor adjustments!
190
Page 27
Calibration:
Work Tour Departure & Arrival Periods
27
Time-of-day choice models should also be compared to survey data to ensure that tours begin
and end at times consistent with observed travel patterns. Typically this calibration is done by
tour purpose.
191
Page 28
28
This shows a calibration summary on work tour duration. The estimated and observed
distributions line up very well, thus it is difficult to distinguish between solid (observed) and
dashed (estimated) lines.
192
Page 29
Model Validation
Check model performance against independent data
sources
Typically traffic counts (by period) and transit boardings
Sensitivity testing
Change network or land-use data
Does model respond appropriately?
Software bug?
Illogical coefficients?
Missing market segments?
Missing variables?
29
Once all of the model system components have been thoroughly calibrated, individually and as a
system, it is time to validate the base-year model. The general approach is very similar to
validating a trip-based model system, but considers more information.
The process of model validation should be thorough. Traffic counts and transit boardings can be
compared against model estimates for an independent assessment of model performance.
Additionally, the model should be tested against various policies to better understand its
suitability for policy\project assessment. The process of validation should inform a careful
adjustment of parameters, variables, or model enhancements to better match observed travel
patterns or ensure logical policy responses. Model validation does not mean adjusting network
speeds to better match traffic volumes!
193
Page 30
15,000
90,000
Obs Entries
Obs Exits
Obs Onboard
Mod Ons
Mod Offs
Mod Onboard
80,000
10,000
70,000
5,000
60,000
(10,000)
(15,000)
50,000
40,000
(5,000)
Colma
Daly City
Balboa Park
Glen Park
24th Street Mission
16th Street Mission
Civic Center
Powell Street
Montgomery Street
Embarcadero
West Oakland
12th Street / Oakland City Center
19th Street Oakland
MacArthur
Ashby
Berkeley
North Berkeley
El Cerrito Plaza
El Cerrito Del Norte
Richmond
Rockridge
Orinda
Laf ayette
Walnut Creek
Pleasant Hill
Concord
North Concord
Pittsburg/Bay Point
Lake Merritt
Fruitvale
Coliseum / Oakland Airport
San Leandro
Bayf air
Hayward
South Hayward
Union City
Fremont
Castro Valley
Dublin/Pleasanton
Daily Volume
200000
30,000
180000
20,000
(20,000)
160000
10,000
(30,000)
Station
140000
Modeled Volume
(25,000)
120000
100000
80000
60000
40000
20000
200000
180000
160000
140000
Observed Volume
120000
100000
80000
60000
40000
20000
0
0
30
Activity-based models can be validated in the same way that trip-based models are validated. As
with trip-based models, we always want to examine our observed data to make sure it is
representative of actual conditions. For example, traffic counts should be from close to the same
time period as each other and to the time of the household survey. Volume and speed profiles
from ITS detector data should also reflect representative conditions. So, days when detectors
were malfunctioning or results were affected be construction projects should not be used.
Traffic volume estimates at a link level should be compared to traffic counts. Since activitybased models typically assign to more time periods, counts by time-of-day are helpful. Transit
boardings can also be compared to observed ridership at a route or route-group level. This slide
shows comparisons between the Metropolitan Transportation Commissions activity-based
model traffic and transit counts to observed data for the San Francisco-Oakland Bay area. The
plot on the upper left shows transit boardings and alightings for BART by station as bars, and
total passengers along the BART route as a line plot.
194
The scatter plot compares estimated versus observed traffic volumes across all count locations on
a daily basis. Similar plots, not shown, were created by time of day. Discrepancies in the
estimated versus observed comparisons suggested further analysis (particularly of network
coding) were necessary. In one case, shortages of volumes lead to the development of a special
generator model for San Francisco airport.
195
Page 31
31
This slide is a list of key calibration and validation measures that have been used by the
Sacramento Council of Governments in evaluating the quality of its activity-based modeling
system. These measures were actually developed when the agency was comparing its new
activity-based modeling system to its trip-based modeling system. As a result, these measures a
quite similar to what the agency would use when validating its trip based model.
196
Page 32
32
Here is another list of validation measures, developed by the Atlanta Regional Commission for
its activity-based modeling system. As you can see, this list goes well beyond the usual tripbased measures to consider measures related to activities and tours.
It should be noted that for most non-work tour-based measures, the only available observed data
is usually household survey results. In order to get a truly independent estimate of these tourbased measures (a source not used in estimation and calibration), it is necessary to either use a
holdout sample of the current survey (which is rarely available), or to apply the model to an older
household survey, combined with a back-casting exercise. If independently gathered GPS data
were available, this might also be used to collect tour-based information.
Knowing that validating an activity-based model will entail a bit more detail, agencies should be
prepared to budget for any extra data collection. Much of the data, however, is from sources
already familiar to agencies from their trip-based model work. The extra effort is mostly in using
more information from existing sources.
197
Page 33
198
Page 34
34
In the second half of todays webinar, were going to focus on the decision of whether or not an
agency should develop an activity-based model and, if so, what questions do they need to
answer.
Activity-based models can take advantage of a much wider range of data than can trip-based
models, and are more complex as a result. They produce a richer dataset for analysis, but that
data requires knowledgeable staff to summarize and interpret it effectively. Custom software,
and sometimes hardware solutions that include distributed computing, can be required to apply
activity-based models. Taking all of this into consideration, you may be wondering whether an
activity-based model is right for your agency.
199
Page 35
35
In this series, we talk a lot about the details of developing and using activity-based modeling
systems. Our hope is that by sharing this information with you, that it will lead to more informed,
better decisions about the development of an activity-based modeling system. Its relatively easy,
however, to get caught up (if not lost) in these details. Taking a big-picture perspective, there are
probably some fundamental questions that an agency should by asking when considering taking
the plunge towards an activity-based model system.
The following list of general questions was developed by an Expert Panel for the Triangle
Regional Model (North Carolina) in November 2011:
What innovative data development and management advances should be considered with
or without an activity-based model?
What is the best organization structure to accomplish regional modeling goals (whether
activity based or advanced trip based)?
In this webinar, we can provide information that will help to answer most of these questions.
Organizational structure is a tough one, though, and something extremely specific to any agency
engaged in travel modeling, with its unique form of governance and relationships to
stakeholders. For example, in the Triangle region (Raleigh-Durham-Chapel Hill), modeling is
provided by a Service Bureau, located at North Carolina State University, and serving two
MPOs, with support from NCDOT. Quite different arrangements exist in other states and
regions. Organizational structure is often developed to address many other issues beyond
modeling. Consequently, we will not attempt to provide answers to questions of the best
organizational structure, other than to note that it is an important consideration.
201
Page 36
36
We can, however, attempt to answer a number of other questions that might help an agency
assess the likelihood of success in developing an activity-based modeling system. This list of
questions was also derived from recommendations made by the Expert Panel in the Triangle
Regional Model review.
Have stakeholders historically been actively engaged in travel demand model analysis?
Has there been a history of resource-based support for travel model development?
Does the agency staff have the required level of technical expertise to run an activitybased model?
Are policy makers or the general public advocating policies that require an activity-based
model?
The answers to these first two questions should be relatively obvious to the agency asking them,
based on first-hand knowledge of the relationships between an agency and its stakeholders and
the history of funding for modeling programs. The answers to these last two questions might be
202
less obvious, because they require familiarity with activity-based modeling systems, which most
persons in the agency wont have had.
203
Page 37
37
TMIP conducts peer reviews around the country for planning agencies who are considering
improvements to their travel demand forecasting models. These peer reviews include a series of
recommendations from the peer review panel to the agency. In four of the last six peer reviews
by TMIP, there was a recommendation to develop activity-based models, based primarily on the
need for increased policy sensitivities that are found in activity-based models. In Arizona, the
panel suggested that more information was needed before a recommendation on activity-based
models could be made (perhaps this webinar series will help to fill that request!). In the Omaha
region, activity-based models were not discussed or recommended. Of the four who
recommended activity-based models, the size ranges from very small (Burlington and Monterey)
to quite large (Philadelphia and Detroit). Philadelphia is proceeding with the development of
their activity-based model and has initiated a consultant contract for this development in 2012,
which will last three years. AMBAG has selected a consultant for their model update, which will
remain as a 4-step model. Burlington and Detroit are more recent and have yet to proceed with
their recommendations.
204
These examples show the range of recommendations and responses to initiating an activity-based
model. Requesting a TMIP peer review can be a measured and thorough means to determine if
you are ready to take the plunge. I am also going to present a list of considerations for whether
an activity-based model is right for your agency.
205
Page 38
Essential
Not necessary
Essential
Desirable
Essential/Desirable
Desirable
Essential/Desirable
Not necessary
38
We have developed a series of checklists so that you can do a self assessment. In this slide,
weve listed skill sets that are particularly useful for an agency engaged in activity-based model
development:
These are skill sets that an agency may already possess, but which are much more important for
activity-based modeling. While consultants may be able to supplement staff expertise, most
agencies find it much more cost-effective to have as much of this expertise as possible in-house.
206
207
Page 39
Use an Activity
Model
Use a Trip
Model
Pricing strategies
Yes
Limited
Non-motorized investments
Yes
No
Yes
No
Transit schedules
Yes, one-way
Yes
Limited
Most
No
Yes
No
Equity evaluations
Yes
No
39
The other question we want to answer is whether policy makers or the public are advocating
policies that require an activity-based modeling approach. In this first session in this series, the
Executive Perspective, we provided several examples of policies that were best addressed
through an activity-based approach. In the second session, Institutional Issues for Managers, we
described how the additional information provided by activity-based models differs from that of
trip-based models and makes activity-based models more appropriate for answering certain types
of questions. This list shown here summarizes these policies contexts:
Pricing strategies
Non-motorized investments
Transit oriented development
Transit schedules
Transit fare policies
Travel demand management programs
Transportation systems management
208
Equity evaluations
209
Page 40
Use an Activity
Model
Use a Trip
Model
Induced Demand
Yes
No
Yes
No
Peak spreading
Yes
No
Yes
No
Yes
No
Yes
No
40
To drill down a bit further, we can imagine a number of performance measures that modelers
would like to be able to predict. In the checklist shown here, weve listed several measures that
activity-based models are designed to predict and for which trip-based models are ill-suited.
These are somewhat independent of specific policy initiatives and are relevant to common
agency work programs, such as long range planning, conformity analysis, and environmental
justice.
210
Page 41
Activity Model
Trip Model
Staff resources
Training needed
Trained
Executive support
Preferred
Assumed
Quality data
Higher expectations
Reasonable
Cost
Updates required
Approach
Transfer or build
Update
Innovation
Lots of possibilities
More difficult
Development
Funding
Work program
41
In the second session, Institutional Issues for Managers, we discussed resource issues relevant to
the development of an activity-based travel demand model. Here, we have listed some of the
resources considerations that an agency should consider when undertaking an activity-based
model development project. These include:
Staff resources
Executive support
Quality data
Cost
Approach
Innovation
Development
Funding
211
Page 42
42
If, and when, you may decide to proceed with developing an activity-based model, there are
some important system design decisions to consider. These include details about the activitybased models, such as what spatial scale should be used, what type of market segmentation you
want to include, what travel behavior choices you want to represent.
These also include details about how the activity-based model will fit into the broader travel
modeling system, such as how will it be integrated with other transportation planning software,
how will other aspects of transportation (freight, long distance travel, visitor travel) be
represented, and how data used in the model will be maintained over time.
212
Page 43
Microzones
Parcels
Spatial Detail
Heterogeneous
Land Uses
Primarily Homogenous
Land Uses
Data Sources
Existing travel
demand model
Aggregate
Measures of
Attractiveness
Aggregate
Partially disaggregate
Disaggregate
Level of Effort
Least
Can be automated
from available data
Most
43
Determining the best spatial scale for an activity-based model is a critical design decision, which
affects most aspects of the modeling system. The spatial scale is not a decision unique to travel
modeling either, since land use modeling, economic modeling, performance measurement, air
quality, operations are all affected by the spatial scale chosen and may indeed have different
spatial scales. One of the bigger benefits of activity-based models is the additional spatial detail
and if traditional traffic analysis zones are used, then these benefits are not achieved. To date,
there are three spatial scales that could be used zones, micro-zones, and parcels. Micro-zones
are approximately the size of census blocks and may be about 10:1 scale with zones. Parcels are
the smallest scale possible and may be several hundred or a thousand to 1 scale.
Data sources will vary depending on which spatial scale you choose, but each level has national
or state data sources that can be used. The spatial scale selected can vary depending on what
function in the model you are evaluating. For example, travel times and costs may be developed
at a zonal scale even though micro-zones or parcels are used for land use data. Some activitybased models have modified the network skimming process to take advantage of micro-zone or
213
parcel level detail for walk to transit, walk or bike modes at the origin and destination end of the
trips and use stop to stop impedances for the line-haul portion of the transit trip. This represents a
hybrid of spatial scales to take advantage of additional detail without bogging down the process
too much. The level of effort is dependent on the spatial scale chosen and the additional effort
involved in developing impedance measures if you use a hybrid approach.
214
Page 44
Auto modes
Drive alone and shared ride modes
Toll and non-toll choices with value-of-time classes
Non-motorized modes
Travel times can be developed from parcel or microzone to network
Networks can reflect elevation, traffic volumes, turn movements, etc.
Activity-Based Modeling: Technical Issues For Managers
44
Another design question for activity-based models is which modes are represented, and at what
spatial scale. The spatial scale is particularly important for walk, bike and walk to transit modes,
since these are short trips and significantly affected by small inaccuracies in travel time or
distance caused by zonal centroids or networks that are missing local streets.
It is preferable to develop transit access and egress from parcels or micro-zones to the transit
stops, but impractical to develop parcel to parcel or micro-zone to micro-zone time matrices, so
hybrid approach to develop transit times as a combination of stop to stop and parcel or microzone to stop works well. Drive access to transit can be separated according to park-and-ride and
kiss-and-ride trips in order to account for parking capacity issues at the parking lots.
Auto modes are commonly separated by occupancy to represent drive alone and shared ride
modes. Additionally, many activity-based models that are used have further segmented auto trips
into toll and non-toll choices and different value of time classes to account for characteristics of
travel that may lead someone to pay (or not pay) a toll for any particular trip.
215
Non-motorized modes can benefit from the additional spatial detail of parcels or micro-zones. In
many cases, highway networks have been supplemented with pedestrian or bike facilities. In
addition, non-motorized networks have added characteristics, such as elevation, traffic volumes,
and turn movements that affect pedestrian and bicycle movements.
216
Page 45
45
The earliest activity-based models had auto ownership and work location models included in
them, but more recent models have benefited from additional long term choices that can be used
in subsequent short term travel choices. Auto ownership models are also expanding to include
the choice of a specific vehicle for each trip and to identify the vehicle type and year for
subsequent use in air quality modeling. Travelers may choose the more energy efficient vehicle
for longer distance trips or the larger vehicle for trips with more passengers and air quality model
results will be affected by these choices.
Transit passes and drivers licenses are strong indicators for someone to choose transit or auto
modes and so are useful choices to include in the mode choice model. In addition, if a person
owns a transit pass, then the price of subsequent transit trips is zero. Parking subsidies also
affects the cost of travel and the likelihood that someone will choose to drive. Parking cost is
already included in most activity-based models and recognizing who bears the cost is important
to include as well.
217
Work and school location choice models are typically included in activity-based models. The
may include the choice of usual work and school choices first, and then whether a traveler will
go to the usual work or school location on the travel day. Some people will stay home or may go
to an outside meeting. The usual work and school location can affect mode choices and
schedules for daily routines, but on some days the person may go to a different location.
218
Page 46
46
This is an example of a model (DaySim) that includes all of these additional long term choices.
In this example, usual work and school locations are estimated first, then work locations for
student workers, then free parking is identified for all workers. Once these initial choices are
made, usual schedule, mode and transit passes are determined for workers, and transit passes are
determined for non-workers. Finally, auto ownership for each household is estimated. All of
these choices are retained as characteristics of the household and persons for further use in
downstream models.
219
Page 47
Model Integration
Should ABM be integrated with other models?
Economic models
Land use models
Dynamic traffic assignment models
Air quality models (EPA MOVES)
Transit benefits (SUMMIT)
47
Activity-based travel demand models do not operate in a vacuum, nor do they answer all the
transportation questions on their own, so integration of activity-based models is expected. The
design question is which type of integration is the most useful. Currently, activity-based models
have been integrated with land use models, economic models, dynamic traffic assignment
models, air quality models like MOVES, and transit benefit models like SUMMIT. These are the
most common models to integrate with an activity-based model, but there may be other types as
well.
Activity-based models are typically also integrated with transportation planning software (like
Cube, TransCAD, EMME, or VISUM) for developing skims and running assignments. This
process involves running the path-building or assignment programs and then reading the matrices
of travel times and costs for use in the activity-based model. The activity-based model, in turn,
will produce new estimates of travel, which are aggregated to zones (or micro-zones) and
assigned to highway and transit networks for assignment. Integration that bypasses input and
output of these files and reads files directly in their native format are faster. In addition, reading
220
skims on-the-fly when you need them in the modeling process, rather than reading and storing
the full matrix, are still being considered to improve the integration of these systems.
221
Page 48
Land Use
Geodatabase
Transit
ITS
Toll Optimization
Activity-based Passenger Travel
Freight Travel
Emissions
Benefit-Cost
48
One example of an integrated system is at the Puget Sound Regional Council in Seattle. PSRC
has integrated a macro-economic model, a land use model, and an activity-based model. In
addition, they have developed a geo-database to store and deliver all the data need for these
models and a series of tools to build alternatives, including a transit service planning tool, an ITS
evaluation tool and a toll optimization model. Once all the models are run, they have linked the
results with a benefit-cost analysis tool and the EPA MOVES model to produce performance
measures for all aspects of the system.
222
Page 49
49
223
Page 50
Type
Airport
External
Visitor
Long distance
Non-resident
Method
Service
50
Activity-based models typically produce resident passenger travel within a region of interest.
Total travel in the region would need to include long distance travel, non-resident travel and
commercial vehicles for the travel demand to be comprehensive. There are many different
methods that can be employed for these types of trips, and they vary widely based on level of
effort, accuracy needed, and the types of policies that may affect these trips. Many activity-based
models are integrated with existing methods to develop these auxiliary trips, but future
consideration can be given to improving the methods adopted for these trips.
224
Page 51
Land uses
Building permits to track changes
Supplemental surveys for government, schools and major
employers
51
Data maintenance and usage, as any modeler knows, can be a time consuming process. Recent
years have seen many improvements in the methods and databases used to maintain and update
data for use in travel demand forecasting processes. The maintenance of networks involves
storing information about possible future projects and combining the lists of projects into
alternatives that can be used to generate a future network. Geo-databases and master networks
have been developed to achieve consistency and accuracy in developing and maintaining
networks. These geographic networks often have more detail than was used in travel demand
networks, but can be brought into the process as more detail is added on the land use side.
Land use data maintenance systems involve tracking building permits to see what development
has occurred against the forecasts. These can be maintained by local governments in charge of
approving building permits. In addition, it is often necessary to conduct supplemental surveys of
government and education employers, as well as major employers with multiple locations, to
ensure accuracy of the employment data.
225
Other data that needs to be maintained for use in activity-based modeling are traffic and transit
counts, as well as bike and pedestrian counts if these are collected. Some regions also conduct
parking inventories, which are quite useful for parking choices in the activity-based model.
226
Page 52
Geodatabase
at Puget
Sound
Regional
Council
52
One example of a geo-database is one that was developed by PSRC for integration of model
networks with the Metropolitan Transportation Plan database and the Transportation
Improvement Plan database. This was built and integrated with editing tools and the ability to
output networks to the travel demand forecasting model. The system was designed so that
planners and non-modelers could code and maintain projects that could then be used directly in
the modeling of alternatives. The geo-database was initially developed to maintain networks and
then was expanded to include land use data.
227
Page 53
53
228
Page 54
54
You may decide that the answers to the questions we have been asking in this second part of the
session lead to the conclusion that an activity-based modeling system might be nice to have, but
perhaps not completely necessary. Nevertheless, your agency might be interested in enhancing
its modeling capabilities. In doing so, there are things that you can do that may better position
your agency to in the future. For example:
229
Page 55
Population Synthesizer
Auto Ownership/Availability (if not already present)
Usual Workplace and School Location Choices
Activity Generation to replace Trip Productions
Destination Choice to replace Gravity Models
University models, visitor models, airport models
55
In addition, there are a number of model components that can be developed for a hybrid model
system. Activity-based model components that can take the place of trip-based model
components include:
Population Synthesizer
Auto Ownership/Availability (if not already present)
Usual Workplace and School Location Choices
Activity Generation to replace Trip Productions
Destination Choice to replace Gravity Models
University models, visitor models, airport models
230
Page 56
1. Population Synthesis
2. Long-term
3. Mobility
4. Daily
Mandatory
Nonmandatory
Home
Individual
Mandatory Tours
4.2.1. Frequency
4.5.1a. Frequency
& Purpose
Year 1 (2009):
Simplified activity-based
travel generation models
estimated, implemented,
and calibrated
At-work sub-tours
4.6.1a. Frequency
5. Tour level
Simplified
Activity-Based
Model for
SANDAG
Trip Distribution
Mode Choice
Trip Assignment
56
One example of a transitional model development effort is in San Diego, who developed a very
simple activity-based model in the first year of their development. These models were fully
estimated, implemented and calibrated for the base year.
231
Page 57
1. Population Synthesis
2. Long-term
3. Mobility
4. Daily
Nonmandatory
Mandatory
Residual time
Individual
Mandatory Tours
Home
Available
time budget
Allocated Tours
Year 2
Individual NonMandatory Tours
4.2.1. Frequency
4.3.1. Frequency
4.4.1. Frequency
4.3.2. Party
4.4.3. Allocation
4.5.1. Frequency
4.2.2. TOD
At-work sub-tours
4.6.3. TOD
5. Tour level
4.3.4. Destination
4.4.2. Destination
4.5.2. Destination
4.3.5. TOD
4.4.3. TOD
4.5.3. TOD
6. Trip level
Year 3
4.3.3. Participation
4.6.1. Frequency
4.6.2. Destination
Year 1
Pop Synthesis & Long-Term Models
Coordinated Daily Activity Pattern
Simplified Tour Frequency Models
4-Step Integration and Testing
Stop Frequency
Stop Location Choice
Trip Mode Choice
Parking Location Choice
Year 4
System Calibration & Validation
Special Market Models
Full PECAS Integration
Sensitivity Testing
57
In subsequent years, the remaining models were completed to supplement this initial
development. This allows for an initial use of the models in the short term (one year) and
subsequent completion of the full activity-based modeling system.
232
Page 58
58
Wed like to conclude the main part of the presentation by reviewing the learning outcomes of
this session. To recap, we hope that you are now better able to:
Discuss the processes used to develop, calibrate, validate, and implement an activitybased model;
Discuss the criteria that agencies should consider when evaluating whether an activitybased model may be right for them; and
Discuss high-level model design decisions that will need to be made when embarking on
activity-based model project, as well as alternative transitional development paths.
233
Page 59
234
Page 60
February 2
February 23
March 15
April 5
April 26
May 17
June 7
June 28
July 19
August 9
August 30
September 20
60
235
Page 61
61
236
time of day models. These adjustments are highly interpretive because the models are not
directly linked to the traffic volumes, but the highway counts offer a second independent source
to inform the model calibration in addition to the household survey data. Calibration refers to the
adjustment of the model coefficients, and validation refers to the comparison of the modeled and
observed outcomes. So we dont split up the data. We are using the same data source in different
ways.
When you talk about adjusting alternative-specific constants, how do you justify or prove the
correctness of these constants?
Joel: Typically, the reason you need to adjust the constants is that the survey data used to
estimate the models is not perfect. The expanded survey data and the modeled population usually
end up with differences. There is some debate about whether the model might actually be more
correct than the survey used to estimate the models. For example, in Atlanta, the survey did not
include group quarters residents living near or on college campuses. In that case, we did not
adjust the modeled trip length frequency distributions. But in other cases, we know that the
model system is not reflecting the true behavior as well because the accessibility estimates are
not perfect compared to the true accessibilities in the region and the survey. This discussion
underscores why we avoid making large, unjustified adjustments to constants, because to do so
would override the estimates of behavior taken from the survey data.
Do you ever compare the activity based model to results from a trip-based model to see if the
activity based model is more accurate or policy-sensitive than the trip-based model?
Maren: Yes, I am aware of three agencies that have performed such comparisons. MORPC in
Columbus did a very thorough comparison using grant money to look at the exact same networks
and inputs using different scenarios in the future, evaluating their sensitivity, and also comparing
the validations in the base year. In the case of validation, both models were very well-validated.
There werent significant differences. For the sensitivity to policies, there were differences. San
Francisco and Sacramento are other regions that have tested differences in policy sensitivity.
They looked at tolling, land use, transit oriented design. Stark differences were found. When
looking at forecasts, some judgment is required in determining what is reasonable, but they were
more able to explain the results from the activity based model, which were more intuitive. The
trip-based model was interpreted to over-predict and under-predict certain changes relative to
what was expected.
It seems multiple days of travel activity would provide a lot more valuable information for
development of an activity based model. Is this happening--are we using multiple-day diaries to
develop activity-based models?
Joel: One of the first activity-based travel surveys was done in Portland in the mid-nineties, and
was a two-day survey. Both days were used, but the days were treated independently. Survey
fatigue is an issue, and trip rates often decrease during the second day. GPS is an option for
238
getting multiple days of data without fatigue. With car-based GPS, the unit can be left in the car
for up to a week. With this method, we get all activities for an entire week, and information
about habitual activities: for example work locations that are visited every day. We can infer
activity locations from the non-habitual activities and get a richer data set, and this is used for
calibration and validation. Research into behavior over multiple days is being performed in
academia. Were not too far out from having models that build behavior in across multiple days.
The advantage would be that certain policies might delay travel but not totally change behavior,
such as license plate allocation or pricing policies.
For calibration and validation, is it more important to match traffic volumes or origin-destination
patterns?
Joel: When we look at assignment results, we consider if were getting the right amount of VMT
for a particular period. Sometimes, a mismatch might be because of issues in the survey data
expansion. Other times the mismatch might be due to a lack or inappropriate assumptions in the
commercial vehicle travel. Screenlines from the traffic volumes can be used to get at origindestination patterns. We can go back to destination choice models and make sure that the districtlevel summaries are consistent with the survey. Perhaps the survey sample was thin for a certain
district movement, but the screenline reveals a more accurate picture, and adjustments might
need to be made to better match the screenline. Then we might go into more detail, for example
looking at traffic by facility type. These are similar procedures as one would follow for
validating a trip-based model, but the focus is on whether there are systematic errors or issues in
the model system that need to be addressed.
239
240
Page 1
Activity-Based Modeling
Session 4: Frameworks and Techniques
April 5, 2012
241
Page 2
Acknowledgments
This presentation was prepared through the collaborative efforts of Resource
Systems Group, Inc. and Parsons Brinckerhoff.
Presenters
John Gliebe, Joel Freedman
Moderator
Maren Outwater
Media Production
Bhargava Sana
Resource Systems Group and Parsons Brinckerhoff have developed these webinars
collaboratively, and we will be presenting each webinar together. Here is a list of the persons
involved in producing todays session.
John Gliebe and Joel Freedman are co-presenters. They were also primarily responsible
for preparing the material presented in this session.
Maren Outwater is the session moderator.
Additional content was provided by Rosella Picado. John Bowman and Mark Bradley
provided reviews of material
Bhargava Sana was responsible for media production, including setting up and managing
the webinar presentation.
242
Page 3
February 2
February 23
March 15
April 5
April 26
May 17
June 7
June 28
July 19
August 9
August 30
September 20
3
For your reference, here is a list of all of the webinars topics and dates that have been planned.
As you can see, we will be presenting a different webinar every three weeks. Three weeks ago,
we covered the third topic in the seriesTechnical Issues for Managers. This session was
designed to help managers understand the process of developing and implementing an activitybased model, some of the issues to consider when evaluating whether to move to an activitybased model, and some of the different development options available. We also provided a highlevel overview of the previous two executive and management webinars.
Todays session is the first of nine technical webinars, where we will cover the details of
activity-based model design and implementation. In todays session, we will prepare participants
for the remaining webinars by explaining choice models, simulation, and the key components of
activity-based models. In three weeks, we will cover population synthesis and household
evolution models.
243
244
Page 4
Learning Outcomes
By the end of this session, you will be able to:
Discuss how household activity-travel diary data is used
to define activities, tours, and daily patterns
Describe how choice model structures are used to
represent key aspects of activity-based model
generation and scheduling
Describe how discrete choice models are used and
applied in activity-based modeling systems
Discuss the various design decisions that are important
to the development of activity-based modeling systems
Activity-Based Modeling: Activity-Based Model Basics
In todays session, we will be covering the basics of activity-based modeling. At the end of this
session you should be able to:
Discuss how household activity-travel diary data is used to define activities, tours, and
daily patterns
Describe how choice model structures are used to represent key aspects of activity-based
model generation and scheduling
Describe how discrete choice models are used and applied in activity-based modeling
systems
Discuss the various design decisions that are important to the development of activitybased modeling systems
245
Page 5
Terminology
Trip
Tour
Day pattern
Schedule
Discrete Choice Model
Monte Carlo simulation
We will first cover some terminology that we will use in this webinar. Some of these terms may
be familiar to you already, but they are worth mentioning so that were all on the same page.
Trip: A movement from an origin to a destination. The trip is the core unit of travel in a travel
demand model (and an activity-based travel model).
Tour: A series of trips that begin and end at an anchor location (typically either home or work).
Day Pattern: A sequence of in-home and out-of-home activities for an entire day.
Schedule: A day pattern with start and end times for tours and/or activities.
Discrete Choice Model: A probabilistic model commonly used to represent the probability of
choosing one alternative from a set of mutually exclusive alternatives.
Monte Carlo simulation: A method of simulating a choice or action, by drawing a random
number from a probability distribution in which the choice is one alternative outcome.
246
Page 6
Key Concepts
Activity-based models attempt to model an entire daily
travel pattern for each individual in a population, as affected
by transportation system level of service
While activity-based modeling systems vary, they all
represent certain key aspects of the activity-travel pattern
creation through integrated model components
Discrete choice models are the most commonly used
analytical formulation for model components and are
applied through Monte Carlo simulation methods
Model design involves developing structural representation
of decision process and how to treat modes, space, time,
and other key model parameters
Activity-Based Modeling: Activity-Based Model Basics
Activity-based models attempt to model an entire daily travel pattern for each individual in a
population, as affected by transportation system level of service. This emphasis on modeling a
representative day for each person is one of the important differences between activity-based
models and trip-based models.
While activity-based modeling system designs vary, they all represent certain key aspects of the
activity-travel pattern creation through integrated model components. We will be describing how
these key aspects are represented in survey data and their representation as model structures.
Discrete choice models are the most commonly used analytical formulation for model
components and are applied in activity-based models through Monte Carlo simulation methods.
Many of you may be quite familiar with discrete choice models, such as the multinomial logit
model that is commonly used in mode choice modeling. We describe how discrete choice models
are used throughout an activity-based modeling system to represent different structures and the
implications for linkages between model components.
247
Model design involves developing structural representation of decision process and how to treat
modes, space, time, and other key aspects of activity generation and scheduling. We will begin
the model design discussion today and will flesh it out in greater detail in subsequent webinars.
248
Page 7
In the first webinar of this series, which was aimed at an executive or managerial-level
audience, we described what an activity-based model was. Since this is the first technicallyoriented webinar on activity-base modeling, it is probably worth discussing here as well.
An activity-based travel model differs from a trip-based model by modeling decisions to
participate in activities. The central focus of the models is whether, when and where to
participate in activities, and for how long. Travel is a derived demand, resulting from the need
for people to engage in activities outside the home. Trips are a means of traveling between
activity locations and decisions related to trip scheduling, such as mode and departure time, are
made to accommodate desired arrival and departure times from activity sites. In advanced
activity-travel modeling systems, these decision are coordinated between members of the same
household. Activity-based travel models are also characterized by their disaggregate
representation of individuals and households, which typically using simulation methods. This
enables us to track these individuals and to effectively use their demographic characteristics in
analysis.
249
Page 8
Arent activities just a fancy name for trip purposes? Not really. Activity modeling does
bear some resemblance to trip-based modeling in the sense that we generate activities, distribute
them to locations, and choose modes for them. Some activity purposessuch as work and school
purposeshave similar labels in the trip-based world. And in fact we do model trips within an
activity-based modeling system. But modeling activities means much more than that.
Activities have a duration (which we model) that has intrinsic value to the participant. People
derive satisfaction from participating in activities, and we assume that the amounts of time that
we observe people participating in activities reflect the utility they derive from them. When we
model the schedule of activities and travel, we take into account the expected amount of time
that an individual would spend in each activity, how they prioritize their time between, work or
school, and shopping and recreational activities, and how much time they devote to travel.
Modeling activities also means allowing for the possibility of in-home substitutions and
tradeoffs, such as telecommuting from home, at-home leisure, eating and other activities. This is
important if were interested in modeling future scenarios in which gasoline prices are higher or
250
if on-line commerce were to become the norm. One response to high travel costs is to undertake
activities at home. In addition, in-home activities of other household members are important. For
example, many parents of young children time their work departure times and forego
discretionary activities so they can be at home for their children.
251
Page 9
Work
Car
Walk
Walk
Car
Daycare
Center
Lunch
Gas
Station
Car
Car
Activity-Based Modeling: Activity-Based Model Basics
Grocery Store
Another key aspect of activity-based travel models is that travel is organized around tours. As
mentioned earlier, tours are a series of trips beginning and ending at home or work anchor
locations. By modeling decisions on a tour basis, there is enforced consistency between the
outbound and return portions of the tour, so that a mode chosen to go to work conditions the
mode available for the return home.
Common to tour-based activity modeling is the identification of a primary destination on each
tour and the insertion of intermediate stops either before or after the primary destination. In
addition, there may be sub-tours within a tour. In this slides illustration, there is a work-based
sub-tour for lunch. This is a contrast from trip-based modeling protocol in which such trips
would be represented as non-home-based tours.
252
Page 10
10
There are a couple of other important themes in activity-based travel modeling, which will
become quite evident in this and other webinars in this series. Activity-based models attempt to
model an individuals entire day of activities and travel. There are various strategies for doing
this, one being the creation of daily activity pattern variables. Daily pattern variables are usually
composed of some combination of tours for various purposes. For example, one pattern may be a
single, simple tour in which someone goes to work and returns home. Other patterns may be
more complex, such as a work tour, with a sub-tour for work-related, activities, and a separate
home-based tour to pick up children from an after-school activity. The important message here is
that tours and trips within the same day are to some degree interdependent. For example, if a
worker has to work late, it may affect other activities in the day. So, perhaps the children have to
find another way home, or a wait a bit longer. In activity-based models, scheduling is subject to
time and space constraints, such that no person can be in more than one place at the same time.
We will cover this in great detail in the ninth webinar in this series and will touch on it briefly
later in this session.
253
Page 11
11
A good way to become familiar with the modeling of activities and travel is to look at data. This
slide depicts sample record from a household activity-travel diary. Each record includes
identifiers for the household, person, survey day, and activity number (sorted). Household
activity-travel surveys will of course collect information on each household and person.
To this we have added information on the persons age, gender and worker status. We have also
appended information from the household records on income and number of autos. Of course, we
could add other variables to these records. In this example, were looking at the activity-travel
records for two persons in the same household for the same day. In this case, we have a 55 yearold male, and a 10 year-old female (father-daughter).
254
Page 12
TAZ
39
82
1290
160
96
39
87
39
39
82
39
87
39
arrive
0
7:05
7:20
15:50
16:20
17:10
19:05
21:10
0
7:05
14:00
19:05
21:10
trip mode
none
auto driver occ 2+
auto driver SOV
auto driver SOV
auto driver SOV
auto driver SOV
auto driver occ 2+
auto driver occ 2+
none
auto passenger
school bus
auto passenger
auto passenger
12
Each record indicates a purpose for each activity, a transportation analysis zone (TAZ) where
each activity took place, arrival and departure times, calculated activity durations and trip travel
times, and a trip mode. As stated earlier, the explicit modeling of activity durations differentiates
activity-based models from trip-based models. This diary example follows a place-based
convention in which there is one activity purpose identified for each out-of-home location.
Oftentimes, people have multiple purposes when going to a particular location, such as a
business meeting that involves lunch, or combining social activities with meals. Nevertheless, it
has become the convention, thus far in activity-based modeling to identify a primary purpose.
This particular example also includes two episodes in which these two household members
interacted with one another. The most obvious one is labeled as joint shop/eat purpose, so
maybe they went to a shopping mall together. You can see that both household members have
the same departure and arrival times. For mode purposes, the older adult was the driver (multiple
occupancy), and the child was identified as a passenger.
255
A less obvious interaction between these two is in the morning commutes. Both persons leave
home at the same time, and the adult escorts the child to school on the way to his workplace.
This drop off event is recorded as escort for the father, and simply as a school activity for
the child, with mode of travel recorded as passenger. After the drop off, we can see the parents
mode switches from multiple to single occupancy vehicle.
256
Page 13
tour mode
none
auto driver occ 2+
auto driver occ 2+
auto driver occ 2+
auto driver occ 2+
auto driver occ 2+
auto driver occ 2+
auto driver occ 2+
none
auto passenger
auto passenger
auto passenger
auto passenger
13
This slide depicts the identification of common tour elements. First, there is the identification of
home-based tours themselves and the stops on each tour. As shown in this example, both persons
have two home-based tours. The first person, the adult, goes to work on the first tour of the day,
while the child goes to school. Both persons participate in a second home-based tour for either
shopping and/or eating out in the evening.
The next step is to identify the primary stop and destination TAZ on each tour. Here, we have
identified the work stop on the adults first tour as the primary purpose and destination. On the
childs first tour, the school stop is the only one, so it is an obvious choice. On the joint tour,
there is also just one stop. Different modeling systems might use a different scheme for
identifying the primary stop on a tour. Typically, the first consideration would be work, school or
college stops. For tours that dont involve work, school or college, such as shopping,
social/recreation and others, a few of different rules have been used. One rule is to choose the
first stop on the tour; another is to choose the stop that has the longest duration; and a third is to
257
choose the stop that is furthest from the tour origin, which will usually be home or the workplace
if a sub-tour. There are pros and cons to adopting each rule.
We also need to identify intermediate stops on tours. In the adults first tour, there is one
intermediate stop (the escort) prior to the primary stop, and two more intermediate stops
afterwards, for household business and shopping.
258
Page 14
day pattern
W-J
W-J
W-J
W-J
W-J
W-J
W-J
W-J
S-J
S-J
S-J
S-J
S-J
14
Earlier we discussed the concept of a daily activity pattern, that relates multiple tours by the
same person in the same day. Daily activity patterns are usually identified by the presence of
tours by types, or even the number of tours by type. It should be noted that stay home all day is
also a legitimate daily pattern, and this happens to be observed quite commonly in diary data,
particularly among certain age groups, the very young and very old. In this example, we have
made up a daily pattern code, symbolized by a letter for each tour type in the pattern. In practice,
different activity-based modeling systems have developed a several different daily activity
pattern coding schemes. Whereas the treatment of tours and stops on tours seems to be quite
similar from on activity-based modeling system to the next, there does tend to be significant
differences in the ways in which daily patterns are coded, which has implications for overall
model system design.
259
Page 15
15
Looking at travel diary data from the standpoint of identifying daily patterns, tours, and various
components of tours, there is a lot of information to consider. This goes well beyond the focus on
individual trips. So, how do we model all of these activities, tours and day patterns?
The first step is to create a synthetic population, representing every person and household in your
study region, with basic demographic information, such as income, gender, worker and student
status, and other attribute data, as needed to support modeling. For each of these synthetic
households and persons, we then predict long-term choices that are important to modeling
activity-travel generation and scheduling. For persons identified as workers, we model the choice
of a usual work place, and for students, a usual school or college location. Auto
availability/ownership is also important and should be modeled as a household-level decision.
It may also be desirable to model what may be referred to as mobility choices, such as transit
pass holding, free parking at work, or participation in a travel demand management (TDM)
program. These are policy-sensitive parameters that we would not expect to be available as
standard inputs from Census level population inputs; therefore, we forecast them, based on
260
models derived from our survey sample, and potentially other sources. Next, we will want to
predict day-level activity patterns and the exact number of tours by purpose.
261
Page 16
16
Having covered, long-term choices, mobility choices, and day-pattern choices, we now have
predictions for exact number of tours by various types. We can then begin filling in the details
for each tour. This would include predicting the primary destination for each tour, followed by
the tour mode and start and end times. Given the primary destination and anchor, we can then
predict whether there will be any intermediate stops on the tour and how many. Given a primary
tour destination and mode and start and end times of day, we can then predict the details of
individual trips on the tours. These details would include individual activity durations, stop
destinations, trip modes and departure times. We do this for every person in our synthetic
population, and we usually process entire households together, particularly in model systems
with explicit intra-household interactions. We then output the results into an activity-trip list. The
trip-list can then transformed into trip tables and assigned to a network.
All of these steps we will discuss in more detail below and throughout the webinar series. What
we have shown here is a representative activity-based modeling processas you shall see,
262
different modeling systems may order the sequence of these steps somewhat differently, and
some systems may describe the prediction steps somewhat differently, or even add steps.
Page 17
5am
?
We assume that persons make many deliberate choices that
collectively result in the activity patterns we observe
Activity-Based Modeling: Activity-Based Model Basics
17
Up to this point, weve mentioned the need to predict many different elements of an activitytravel pattern, but we havent talked about the methods for doing so. In this next section, we will
touch lightly on the methods of predicting travel choices. We begin from the theoretical premise
that these choices are not simply random, but rather the outcome of deliberate decision making.
Further, we assume that elemental choices that we try to model collectively represent the
activity-travel patterns that we observe.
263
Page 18
Destination (TAZ#)
2 . 235 236
Starting Time
5am
18
We have talked about several different choice dimensions. Anyone who has worked with travel
demand models should already be familiar with mode choice models. In this diagram, we depict
the choice from among five different mode choice options. However, we can model other types
of choices, including the choice of destination and time of day. In these two examples, were
considering choices made in space and time respectively, which are really continuous
dimensions. For analytical convenience, we can parse them into discrete units that make our job
considerably easier. The fundamental unit of analysis might be a zone, or it could be something
even smaller, such as a grid cell, micro-zone or even parcel. Likewise, we can parse time into
intervals and choose a starting time interval for our activity. Activity-based modeling systems in
use to date have used 60, 30 and even 15-minute decision intervals.
The ways in which space and time are transformed into discrete intervals has important
implications for how these data are processed in surveys, for the creation of networks loading
points, assignment time intervals, and the maintenance of geo-databases. Well discuss the data
implications of model design in more detail later.
264
265
Page 19
Workplace
Location (TAZ#)
3+
Buy Transit
Pass?
2 . 235 236
No
Yes
3+
Add Stop
Before Primary
Destination?
No
Yes
Add Stop
After Primary
Destination?
No
Yes
19
In addition to different units of analysis, our models will need to consider different time
horizons. In this illustration we have listed a few of the common long-term choice models that
appear in activity-based modeling systems. These include the choice of workplace location
(defined as a TAZ), the number of household auto to own, and whether an individual would
purchase a transit pass. While workplace and transit pass are individual choices, the number of
autos to own is an example of a household level choice. As you might imagine, in real life, these
three decisions might be somewhat interdependent. The number of cars owned might depend on
where individual household members work. In the case of some part-time workers, however, the
direction of causality might be reversed. In addition, whether someone bought a transit pass
might also depend on workplace and the availability of autos. The sequence in which these
decisions are represented in activity-based modeling systems is part of the model design.
This next group of choices represents model aimed at daily tour pattern generation and certain
stop details. Here, you may notice that we are actually predicting the number of shopping tours
in a daily pattern, given the existence of at least one. Then we have a couple of models that
266
predict whether to add an intermediate stop before or after the primary destination stop. So, we
can use a choice mode approach to predict the frequency of occurrences of something like tours,
where the number is likely to be small (say 0, 1, 2, 3+). We can also use a binary choice structure
to predict a yes-no type of response. The add stop model might be applied multiple times. For
example, applied once to predict an initial insertion of an intermediate stop, then re-evaluated to
predict whether there is room in the schedule for more. As you might imagine, there are other
ways that we could represent this decision process, including linked and ordered choices. In
practice, we have found that some of the simplest model structures work the best over a wide
range of input cases.
267
Page 20
Conditional choice
Tour Destination
(TAZ#)
1
Tour Mode
2 . 235 236
Tour Mode
SOV Car Walk Bus LRT
Pool
(Upstream conditioning
other decisions in between)
Tour Start & End Times
Intermediate Stops
Stop Destinations
Tour Starting
Times
5am
Tour Ending
Times
7am 8am . 10pm 11pm
Trip Mode
20
Because some many choices are interdependent, activity-based modeling system designs try to
capture these interdependencies to the extent practical. Many people believe that what we have
described as separate choices, such as mode and destination, are really bundled choices; for
example, choosing between combinations of mode and destination. This first diagram represents
the joint choice of tour primary destination and mode in a hierarchical manner. We could
enumerate every combination of destination and mode, all on the same level, but that is not
necessarily more accurate and is definitely less practical in terms of model estimation and
application. In this example, we have chosen to represent the choice of destination first, and
conditional upon destination, we have the choice of mode. The important take away is that the
choice of destination conditions the choice of mode, and that the composite travel times and
costs of the modes available to travel to each destination alternative affect the choice of the
destination.
Another example of a joint or conditional choice would be the choice of tour starting and ending
times. Here there is an obvious logical constraint being enforced in which the tour ending time
268
intervals must be later than tour starting time intervals. Implicit in this choice of starting and
ending times is the tour duration. The choice hierarchy seems clear due to temporal ordering of
starting times before ending times; however, the utility of ending time and duration may
influence the tour starting time.
Yet another example of a conditional choice would be the choice of trip mode, conditional upon
tour mode. This would seem to be a rather obvious hierarchical relationship in which the mode
chosen for the whole tour dictates what is available for individual trips on the tour. In this
example, the person chose to walk for the tour mode, leaving SOV and bike unavailable for the
subsequent trip mode choices. In an activity-base modeling system, however, there may be one
or several other choice decisions that take place in between tour and trip mode choice. For
example, after choosing the tour mode, there may be the choices of tour start and end times, the
decision of whether and how many intermediate stops to insert, and the choices of destinations
for those stops. So, while the choice of trip mode is certainly conditional upon tour mode, it is
also conditional upon a handful of other choices that take place upstream.
269
Page 21
Mandatory + Joint
Non-Mandatory
Mandatory
Stay Home
1+ Escort Tour
Daily Activity
Pattern (Model B)
1 +Shopping Tour
1+ Work Tour
Stay Home
Daily Activity
Pattern (Model A)
21
At this point, it should become apparent that there are many different ways in which these choice
elements can be represented and integrated into an activity-based modeling system. Here is
another example, depicting the representation of the choice of an overarching day pattern for an
individual.
The first model (Model A) represents day patterns as combination of tour types. Given a large
number of combinations of tours of different types, there could literally be thousands of
individually defined alternatives. Although practically, this type of day pattern model, would
eliminate those that are observed rarely and group certain alternatives. The exact number of tours
of each type would be chosen in a subsequent series of models.
The second model depicted here defines day pattern alternatives differently by characterizing the
day patterns as being either mandatory or non-mandatory, with a secondary choice of whether
to include joint activities with other household members. Mandatory is defined as a pattern
involving work, school or college activities, but may include other, discretionary activities, such
as eating out, shopping, and social/recreational. A non-mandatory day pattern would include
270
only discretionary activities. Joint activities with other household members are an extra
dimension that could be added to either a mandatory or non-mandatory day pattern. The exact
number of mandatory, non-mandatory activities and tours, as well as joint activity participation
would be determined in downstream models. What is shown here is a simplification.
Again, these are model design decisions. We will begin to discuss different approaches to model
design in the second half of this webinar. We will get into the finer points of model design
options in the next several webinars in this series. For example, Webinar 8 is devoted to activity
pattern generation.
271
Page 22
Choice Theory
Many decision rules and theories out there
Lexicographic ordering, min-max ranking
Elimination by aspects
Risk minimization strategies, prospect theory
22
We have discussed the various ways in which activity-travel choices are represented structurally.
Next, we will discuss the mathematical formulations for these choice models. There are a
number of competing rules and theories that have been proposed by behavioral psychologists and
economists to describe the ways in which people make choices. Some of these include
lexicographic ordering and ranking strategies, elimination by aspects, risk minimization
strategies, and prospect theory. These all have merit in describing certain decision makers in
certain contexts.
In activity-based travel modeling, we tend to use utility maximization as a theoretical
underpinning. The assumption is that people choose the alternative that provides them with the
highest utility among available alternatives. This has been found to be robust over a wide range
of decision makers and choice contexts. While it carries with it certain assumptions, it is applied
probabilistically in model formulations, which allows us to account for measurement error and
random heterogeneity in the population. Some of its somewhat less realistic assumptions include
272
that the decision maker has full knowledge of the attributes of each alternative and pays equal
attention to all available alternatives.
273
Page 23
Missing variables
Unobserved taste variation
Measurement error
Incorrect functional form
23
Many of you have been exposed to discrete choice models either through your academic training
or on-the-job experience in travel demand modeling, although I am sure there are a few persons
in the audience who less familiar. Most travel demand modeling professionals are at least
familiar with mode choice models. So, this is not intended to be an extensive tutorial on discrete
choice models. Our purpose here is to highlight certain important aspects of discrete choice
models that are central to their use in activity-based travel modeling. We want to make sure that
you are familiar with important terminology. In particular, we want to discuss the roles of choice
sets, composite utility or log sums, and how models are applied in a simulation environment.
Starting with Random Utility Theory, we assume that the decision-maker selects alternative that
is perceived to offer the maximum utility from a set of alternatives that are mutually exclusive,
which we call the choice set. The observer does not know utilities; however, they may be
inferred from the choices made.
Sources of error include: missing variables, unobserved taste variation (preferences),
measurement error (actual versus perceived travel time), and using the incorrect functional form
274
(linear, non-linear, hierarchical, etc.). We treat these errors and random and additive, resulting in
the utility formulation shown here. Total utility is composed of a systematic portion Vj, which
we represent through the variables in our model, and a random component symbolized by the
epsilon error term.
275
Page 24
Choice Probabilities
Probability of choosing alternative i from a set of choice
alternatives C
P (i : C ) Prob U i U j , j C
Prob Vi i V j j , j C
exp V
j
exp(V j )
exp(Vi V j )
24
276
Page 25
+
+
+
+
a * in-vehicle time
b * fare
c * access time + egress time
d * wait time
mode-specific constant
25
Lets use a mode choice example to illustrate how a utility function is formulated. The utility
equation for transit is shown on this slide. Utility equals the weighted sum of the attributes of the
alternative. The weights in the model are known as model parameters, shown here as a, b, c, and
d. These parameters can be estimated from survey data, borrowed from another model, or
asserted based on experience. The parameters convert the modal attributes in various units such
as minutes and cents to a general value called a utile (since they measure utility). This has
important implications for how the weights can be compared to one another. Note that there is
also a term called a mode-specific (or alternative-specific) constant. This represents the value (in
utiles) of all of the attributes of the alternative that are not explicitly listed in the utility equation.
In the case of transit, this could include difficult-to-measure factors such as transit reliability,
transit safety, and the influence of weather on the choice of transit.
277
Page 26
2 . 235 236
jn
P(i )
exp(V j|n q n )
jn
exp Vm q m ln exp(V j|m q m )
m
jm
26
Earlier, we were discussing several examples of joint and conditional choices. Activity-based
modeling systems make extensive use of such hierarchical or nested choices. For example, we
looked at the nested choice of model conditional upon destination. We also considered the nested
choice of tour ending time, conditional upon tour starting time. We can represent the conditional
probability of a choice that appears in a lower-level nest upon the choice made in the upper-level
nest as follows in these formulas. The theta parameters are dispersion terms that reflect the
correlation between alternatives in the same nest. In order to be consistent with utility
maximization parameters, theta must have values greater than zero and less than or equal to one.
The term highlighted here represents the composite utility of the nested alternative. Notice that
the denominator for the lower-level choice (mode) appears in the utility expression of the upperlevel choice of destination zone. Because we take the natural log of this sum, this term is
commonly referred to as the log sum. It represents the maximum expected utility that may be
derived from the lower-level choice, which in this case is mode. In the choice of a destination
shown here, the log-sum term represents the mode-weighted accessibility for travel to each zone
278
alternative. Another portion of the utility of the zone alternatives Vm is shown here, and
represents other attributes of the zone, such as attraction variables.
Thus, it is common in activity-based models to use composite accessibilities, such as mode
choice log-sums to account for travel times and costs by all available modes when choosing a
destination. The assumption is of course that the destination is chosen first. We will talk more
about how log-sums are used in activity-travel model components in the second half of this
webinar and in future webinars.
279
Page 27
2 . 235 236
27
The choice set is the group of alternatives considered to be available to the chooser in a given
choice context. The roles of choice set formation and restrictions are important in activity-based
modeling systems. This is particularly true for nested choices and conditional choice
relationships, as discussed above. In conditional choice contexts, the upstream model choice will
in many cases condition the availability of alternatives downstream. For examples, as we showed
earlier, the choice of tour mode conditions the availability of certain trip modes. Generally
speaking, if a person does not choose to drive for the tour mode, then we would not expect drive
to be available for any trip on the tour. The same would be true with bicycle.
In addition, the presence or absence of an alternative in a lower-level choice may greatly affect
the composite utility of the upper level choice. So, in a policy context, if we were to add a new
transit service to the region that would greatly improve travel time by certain zones, then this
addition of a new alternative to serve those zone pairs would make those destinations more
attractive. This change in accessibility would be reflected in the mode choice log sums that
280
would be used by an upstream destination choice model and possibly even long-term workplace
and auto ownership choice models.
281
Page 28
Activity/Tour-Based/SimulationMode Choice
1. Predict probability of each simulated chooser selecting each
mode for a specific O-D pair and purpose.
2. Use Monte Carlo random draws to predict mode choice.
3. Sum over choosers and purposes, grouped by O-D pair, to
form trip tables for network assignment.
Activity-Based Modeling: Activity-Based Model Basics
28
282
Page 29
2.
3.
SOV
0.56
0.56
HOV
0.28
0.84
Bus
0.03
0.87
LRT
0.08
0.95
Walk
0.01
0.96
Bike
0.04
1.00
SOV
0.00
0.56
HOV
0.57
0.84
Bus
0.85
0.87
LRT
0.88
0.95
Walk
0.96
0.96
Bike
0.97
1.00
29
There are three basic steps in Monte Carlo prediction. First, predict the probability of each
choice for each household or person making a choice. Here we use mode choice to illustrate an
example, but the same applies to any of the choice models we have discussed in this webinar.
Next, calculate the cumulative probability of the array of choices, as shown here, such that they
add up to 1.0. These values represent the upper bound of prediction bins.
The second step is to draw a random number from a uniform distribution on the unit interval.
The third step is to select the range on the cumulative probability array that includes the random
draw. In this particular example, we drew .76, which falls into the bin range for the second
alternative, Auto HOV. Had we drawn a different number, say 0.33, wed be choosing Auto
SOV.
283
Page 30
30
Monte Carlo simulation has advantages and disadvantages compared to expected values used in
trip-based models. The key advantage of Monte Carlo simulation is that explanatory variables
can be included in models with little computational overhead (as opposed to aggregate models,
in which each market segment increases the number of calculations exponentially). Monte Carlo
simulation also can provide a distribution of results, though in practice this has not been fully
taken advantage of. The main reason for not taking advantage of the ability to forecast a
distribution of outcomes is that it would require excessive total run time to do, say, 100 runs.
Monte Carlo simulation also helps us avoid aggregation bias in prediction.
284
Page 31
31
This slide is illustrates the issue of aggregation bias using logit models. On the horizontal axis
you have the cost of a choice for two different decision makers, A and B. Follow the red lines.
Note that the probabilities for each individual, as predicted by the model are quite different.
Now, if we were to aggregate these two individuals and take their average cost, wed obtain a
different probability that is somewhere in between. This is shown by the blue lines.
However, notice that the probability of this average cost is different than the average probability
we obtain when we calculate each persons probability
285
Page 32
32
This slide illustrates another aspect of aggregation bias. The average impact of a change is not
equal to the impact calculated at the average of the explanatory variables. This is symbolized by
the tangents to the curve, representing slopes at each point. Again, the red lines represent the
individual outcomes, and the blue line represents the slope corresponding to the averaged
outcome. Due to the sigmoid (S-shape of the curve), the logit model is most sensitive (elastic) to
change in inputs at its center region, and is relatively less sensitive (inelastic) to changes in
inputs at its top and bottom ends. This is one reason why some aggregate models predict larger
shifts in response to scenario inputs changes than disaggregate models.
An example of this might be a mode shift in response to a new toll charge. Imagine the perceived
cost of the toll being affected by personal values of time, where Person A has a high willingness
to pay (so perceived cost is not so onerous) and Person B has a low willingness to pay (so
perceived cost is considered to be very onerous). Because both persons are already at the far ends
of the distribution, they are less likely to react to a cost change by changing their baseline
choices. By grouping travelers under a single average value of time, however, the perceived cost
286
represents an average condition, the blue slope, and has the potential to overestimate the
elasticity of response to the toll.
287
Page 33
Ways to compensate
Fix seed
Average results
Sample replication and weighting
33
The key disadvantage of Monte Carlo simulation is that multiple runs are required in order to
determine the expected values, or average results, for certain model outputs. This has
implications for forecasting, but most practical activity-based models compensate for this.
The amount of variability in model results is dependent on the number of decision makers in the
choice decision, and the size of the probability of the choice. For example, lower probability
choices have more variability in their outcomes than higher probability choices. Because most
outputs from activity-based models are aggregations of choices, one run of the model can be a
sufficient indication of the expected outcome from a policy. For example, regional VMT, VHT,
district-level tour flows, tours and trips by mode, and higher facility-type link estimates and
transit line boardings are very stable from run to run. However, more disaggregate analysis, such
as TAZ-level origins and destinations, lower facility type link loadings, and lower ridership
transit routes, can have more variation and therefore multiple runs of the model system may be
required, where results are averaged across the runs.
288
One way to compensate for Monte Carlo variability is to fix the random number seed, such that
the results will only vary according to changes in inputs. This ensures stability from run to run,
but at the cost of representing only one possible outcome. Also, results can be averaged to obtain
an expected value. Finally, the synthetic population can be replicated, and the model can be run
once where each sample can be weighted to obtain an average or expected outcome.
289
Page 34
290
Page 35
35
Having examined how activity modeling choices are represented in both data and model
structures, we can now talk about model system design. If you study the different model designs
in detail, you will find different philosophies expressed in terms of how to define daily activity
patterns, how activities themselves are defined, and the sequencing of model steps. In addition,
there are several key design elements that must be decided upon. These include:
291
In addition, we need to decide on the sequencing of model steps and how the model system will
be integrated with network supply models.
292
Page 36
Person Type
Age
Work Status
School Status
Full-time worker
18+
Full-time
None
Part-time worker
18+
Part-time
None
Non-working adult
18 64
Unemployed
None
Non-working senior
65+
Unemployed
None
College student
18+
Any
College +
16-17
Any
Pre-college
Non-driving student
6 16
None
Pre-college
Pre-school
0-5
None
None
From
San Diego
ABM
36
One of the first design elements to be considered is the synthetic population and the role that
person and household attributes will play in the overall model design. Although we have been
telling you throughout this series how beneficial it is to model persons in disaggregate form, it is
often useful to create person-type categories for the sake of modeling convenience. The table
shows an example of person-type categories and definitions. Person-type categories can be used
for a number of purposes:
As a basic segmentation for certain models, such as daily activity pattern models.
To summarize and compare observed versus estimated data and calibrate models.
As explanatory variables in models.
As constraints on alternatives that are available; for example, mandatory activities are
only available to workers and students.
Some categories, such as work and school status, are there because they make logical sense.
Determining the proper cutoff points for categorical variables related to age and income are
293
usually derived in the model development process empirically through descriptive statistical
work. For example, a community with a large population of retirees might use different age
groupings than a college town. Not having proper cutoff points could result in important market
segments being under-represented in model specifications and estimated models not explaining
as much behavioral variation in the population as they could.
294
Page 37
HH by Income
HH by Size
SF AGE_HH
0
24
1
23
1
43
0
32
1
34
0
49
1
67
1
15
0
12
Household File
HID SIZ INC WRK
1
1
1
1
3
1
1
0
5
1
1
1
SF
0
1
1
AGE_HH
24
43
34
Person File
HID PID AUT INC WRK GEN AGE EMP
1
1
1
1
1
0
24 1
3
1
3
1
0
0
43 1
5
1
0
1
1
1
34 0
37
The attributes and attribute levels specified in the synthetic population should be consistent with
the person types actually needed in the model. This slide shows one possible approach for
construction of a synthetic population. In the first step, control totals for each zone are generated.
In the base-year, these can be obtained from Census. In the future year, these control totals can
be based upon growth factor models, allocation methods, and/or land-use models. In the second
step, the aggregate characteristics of the synthetic population are defined, in the form of a joint
distribution of household attributes for each zone. In the third step, individual person and
household records are drawn at random from a sample of households, such that combined they
meet the joint distribution defined for the zone. There are other approaches, such as weighting
disaggregate PUMS records directly according to the control totals, at both a household and a
person level, and then selecting households according to those weights.
These techniques will be described in more detail in the Webinar 5. Some of the decisions that
need to be made will include which household and person attributes are most important to
295
control and at what level of spatial resolution, given the available data, and what other household
or person attributes can be added as uncontrolled attributes.
296
Page 38
Long-Term
Choice
Models
Medium-Term
Mobility
Choice
Models
Transponder Ownership
Activity-Based Modeling: Activity-Based Model Basics
38
Long-term choice models include models of regular work location for workers, and school
location choice for students. These have become fairly standard in all activity-based modeling
systems. Possible medium-term mobility models include auto ownership, transit pass holding,
free parking reimbursement (for workers who work in parking-constrained areas such as
downtowns), and transponder ownership. The reason that these models are placed first in the
model chain is that these choices are made on a long-term temporal scale, and they condition
later choices that are made on a daily or even hourly time scale. Some model systems provide the
ability to turn off these models, keeping their choices fixed from a baseline model run, so that
only short-term or daily decisions can be analyzed. One motivation for this feature is the ability
to measure ramp-up periods. For example when a new toll road is first opened, workers have
not yet adjusted their work location choice based on the level-of-service offered by the facility,
and many households have not acquired transponder units. In the longer-term, workers might be
expected to adjust to the level-of-service provided by the facility and households who are willing
to pay for the toll facility would be expected to acquire transponder units. In webinar 7, we will
297
cover both long-term choice models and medium-term mobility models and discuss these issues
in more detail.
298
Page 39
Treatment of Space
Spatial Representation
Diagram
Zones
Already exists for most MPOs
The most aggregation error, particularly for non-motorized and transit modes
Sub-zones
Created by buffering around transit lines, stops
Improved representation of walk-transit, but may not be consistent with skims
Doesnt help with non-motorized representation (intra-zonal walk and bike)
Micro-zones
Created by sub-dividing zones (7-10:1)
Best representation of transit accessibility when coupled with stop-stop skims
Improved representation of non-motorized time
Parcels
Created via parcel database
Improves representation of walk-transit, but need to make consistent w/ skims
Best representation of non-motorized time
39
As discussed earlier, another important design element is the treatment of space and the
calculation of accessibilities. This slide shows commonly-used spatial systems in activity-based
models, and describes some trade-offs between them. There are generally four different
treatments of space that can and have been used in activity-based models.
Transportation analysis zones, or TAZs, were used in earlier versions of models, and are still
used for smaller regions where TAZs tend to be quite small. The advantages of TAZs are that
they are readily available and it is generally easy to estimate land-uses at the zone level. The
disadvantages are that they typically have some level of aggregation bias with respect to intrazonal and close-in travel, particularly for transit access/egress and non-motorized travel.
Another approach involves developing sub-zones by buffering around transit stops, and using the
sub-zone definition to over-ride skimmed walk access/egress times to/from transit. This approach
is also commonly used in trip-based models. Advantages of this approach are that it is easy to
create the buffers using simple GIS procedures, and it offers an improved representation of
299
transit access/egress. However, there is still some aggregation bias with respect to transit skims,
and the approach is not helpful for non-motorized travel.
Spatial disaggregation, to the extent that it can be done realistically, improves the ability to
measure activity attractiveness and travel impedance. And differences in attractiveness and
impedance, which vary a LOT within traditional zones, have VERY LARGE impacts on people's
travel choices. As activity-based models become spatially more disaggregate, it is becoming
more and more feasible to realistically incorporate non-motorized modes into the models.
The third approach is to code a set of micro-zones by sub-dividing zones, using grids, or census
blocks. Advantages include the ease of creating micro-zones, and the approach offers a very
precise measurement of walk access/egress to/from transit, especially when coupled with stoplevel transit skims. However, employment data can be more difficult to allocate to micro-zones
than to TAZs, depending on the availability of good employment data.
Finally, parcels can be used in activity-based models. They are sometimes available from
existing sources, and are the most disaggregate with respect to non-motorized distances and walk
times. However, parcels are not necessarily stable across time, and allocating employment data
to parcels can be challenging. We will cover this topic in even more detail in Webinar 6.
300
Page 40
Accessibilities
Synthetic
Population
Downward
Integrity:
Choices made
in higher
models affect
choices made
in lower
models
Mobility
Choices
Model
Inputs
Long-Term
Choices
Daily
Activity
Patterns
Trip
Assignment
Upward
Integrity:
Expected
utility of
making
choices in
lower models
affect choices
made in higher
models
Model
Outputs
Activity-Based Modeling: Activity-Based Model Basics
40
Webinar 6 will also discuss the calculation and use of accessibilities in the activity-based model
system. While activity-based models can vary in structure, this diagram shows the location of
tour and trip detail choices (tour mode, primary destination, intermediate stop location and trip
mode) in a typical model stream. As the model system progresses, travelers make decisions:
whether to travel, where to go, how many stops to make, what mode to choose, and so on. Earlier
decisions influence and constrain the decisions made later; for example, the number of vehicles
owned, as modeled in the auto ownership (mobility) model, influences the number of tours and
the mode used on each tour. The mode used for the tour then influences the location of stops on
the tour, and so on. This is referred to as downward vertical integrity.
Activity-based models also use information from models that are lower in the model chain to
inform the choices made by decision-makers in upper-level models. This information typically
takes the form of accessibilities that are based upon all of the information that is relevant for a
lower level choice. For example, a mode choice log-sum, which reflects accessibility by all
301
modes of transport, can be used to inform the choice of destination for the tour or stop. This is
referred to as upward vertical integrity.
302
Page 41
41
Log-sums can be plotted. This a map of the mode choice log-sum from all micro-zones to
downtown San Diego, specifically for travelers who reside in households with at least as many
cars as adult drivers. The darker colors indicate more accessible zones and the lighter colors
indicate less accessible zones from downtown, consider the time and cost of travel by all modes
of travel, but specifically for auto sufficient households. Log-sum variables may be created for
specific market segments and for mode choice are typically segmented by household auto
availability or sufficiency (which relates workers or drivers to available autos). Market
segmentation of log-sums is yet another key design element.
303
Page 42
Zone 3
Zone 4
Origin &
Primary
Destination
Primary
Destination
Work Tour
Origin
Zone 2
Intermediate Stop
Tour Data
HH #
Per #
Tour
#
Purpose
Origin
TAZ
Destin.
TAZ
Outbound
Stop1 TAZ
Return
Stop1 TAZ
Mode
Outbound
Time
Return
Time
1023
Work
Transit
7 :30AM
5:00 PM
1023
Work-Based
Walk
12:00 PM
1:00 PM
42
At beginning of this webinar, we looked at some household diary data from a survey. Here is a
graphic depiction of a days tour pattern for one person, with a table that combines all of the
activities into two tour records. A worker leaves home in the morning and travels to work in zone
3. The worker then goes to lunch in zone 4 and returns back to work. At the end of the day, the
worker returns home, but stops on the way in zone 2 for some groceries.
The definition of tours and components of tours is fundamental to activity-based modeling. This
travel pattern consists of two tours. One tour is a work tour which consists of the set of trips from
home (zone 1) to work (zone 3) to shop (zone 2) back to home (zone 1). In this example, the
home location is the tour anchor or origin (the start and end of the sequence of trips for the
tour). The primary destination is the key location on the tour that defines the tour purpose. It is
often the main reason for travel. For a work tour, it is the workplace.
How to determine which stop in the tours is the primary destination is one key design decision.
While it is possible in recent tour-based household surveys to ask the primary purpose of the
tour, this has not always been the case and is certainly not true in all surveys, particularly older
304
ones. Using a hierarchical typology based on activity purposes is one method, which works well
for work, school and college purposes, but for other purposes primacy is less clear. Other tiebreaking rules include the first stop on the tour, the stop furthest from the home anchor point,
and the stop with the longest duration. This has important implications for the construction of
tour schedules, since time window availability criteria for the insertion of intermediate stops
would be influenced by both activity duration and travel time to the primary destination.
There can be zero or more intermediate stops on the tour, which are stops made between the
anchor location and the primary destination. Some activity-based modeling systems refer to
sequence of one or more stops between the anchor location and the primary destination as the
first half of the tour, or outbound half; and the sequence of one or more stops between the
primary destination and the anchor location as the second half of the tour, or return half. On this
tour, there is one intermediate stop on the return half of the tour, between work and home. There
are no stops on the outbound half, which is between home and work. Whether to model stops on
tours using this half tour schema, or a more sequential method, is another design decision. The
sequence of trips between work and lunch is referred to as a work-based tour (or sub-tour). In
this case, the anchor location for the tour is the workplace, and the primary destination is lunch.
There are no intermediate stops on this tour. Another design decisions is whether to allow nonwork locations to be anchors for sub-tours, such as school.
305
Page 43
Activity Purposes
Mandatory activities
Maintenance activities
Escort, Shop, Other Maintenance (e.g., personal business)
Activities performed on behalf of household
Discretionary activities
Eating out, Social/recreation, Other Discretionary (e.g., medical)
Most flexible in terms of generation, scheduling
43
In general, disaggregation of travel purposes by activity types makes activity-based models more
sensitive to variations in travel behavior than trip-based models and allows them to be more
accurate when matching person types with activity locations and times of day.
Activities are often grouped into three key categories according to priority in the daily activity
pattern schedule: Mandatory, Maintenance, and Discretionary. Mandatory activities consist of
work and school. They are the least flexible in terms of generation and scheduling, and are the
basic building blocks of activity schedules for workers and students. Some model systems
differentiate between work-at-home (telecommuting) and work-out-of-home activities. Some
models also categorize school activities by grade level.
Maintenance activities include escort, shopping, and other maintenance which can include
doctors visits. Some modeling systems model certain purposes explicitly, while others combine
them into more general categories, like other. This is a design decision that should depend on
local modeling needs. For example, in areas with a large contingent of senior citizens, explicit
modeling of a medical activity purpose may be desirable.
306
Many of these activities are performed on behalf of the household, such as picking up or
dropping off household members, or going grocery shopping. In model systems that represent
joint travel explicitly, the escort purpose may be replaced by more detailed descriptions, as we
shall see in the next slide. Finally, discretionary activities include eating out, visiting, and other
recreational activities. They are the most flexible in terms of generation and scheduling, and
often substituted for in-home activities, particularly for households with poor accessibilities to
recreational opportunities. In addition, some modeling systems differentiate activities on workbased sub-tours from those belonging to the main home-based tour. One reason for this is
because sub-tours tend to be more constrained in terms of time; therefore, activities on workbased sub-tours are likely to have significantly shorter average durations and travel distances.
307
Page 44
Joint Travel
Partially joint travel: Person 1 (worker) escorts person 2 (student)
to school
Zone 1
Zone 4
Zone 3
Zone 5
Person 2: student
44
Joint travel refers to tours and trips that consist of at least two persons from the same household.
Research shows that over 80% of shared-ride trips are made by members of the same household.
Depending on the model system design, it may be represented implicitly or explicitly. Explicit
representation of joint travel is useful for modeling high-occupancy vehicle demand, mode
choice elasticity, and vehicle allocation among household members. There are two types of joint
travel. Partially joint travel refers to tours where picking up or dropping off passengers occurs
on the tour, as shown above. Fully joint travel refers to tours where all members travel to all
stops on the tour together, shown below. We will discuss different ways to generate joint travel
in Webinar 8, and will revisit joint travel in Webinar 10, where we discuss tour and trip mode
choice.
308
Page 45
Treatment of Time
Different temporal
systems used
5 time periods, hourly,
half-hourly, continuous
45
Disaggregated treatment of time in activity-based models makes them more sensitive than tripbased models to changes in level of service as well as the idiosyncratic preferences of individuals
doing things at certain times of day (such as meals) and business hours. This makes many
activity-based models a more effective tool for modeling peak spreading behavior, congestion
pricing, and TDM policies that aimed at shifting travel demand across time.
Although real time is continuous, many activity-based models often treat time as intervals, or
time periods. As we saw earlier in this presentation, we can parse time into intervals in a way
that allows us to choose starting and ending times for activities using discrete choice models.
These intervals range from half-hourly, to more aggregate. Sometimes multiple temporal systems
are used in the same model; for example, network skims are created for 4-5 time periods but
activities are scheduled into half-hourly intervals (or even assigned continuous start and end
times). Some activity models schedule activities in quasi-continuous time, either by using
continuous functions to predict activity duration, given a starting time. Moving towards finer
temporal resolution in both activity-decisions and network assignment is an area of growing
309
research. In Webinar 9, we will cover the treatment of time in activity-based models, and ways to
schedule tours, trips, and activities.
310
Page 46
Activity Scheduling
1. Schedule Work Tour
2. Calculate residual time windows
5:00
2-Disc
>75:00
9 P.M.
23:00
46
In addition to decisions regarding temporal resolution, one of the other fundamental design
decisions in activity-based modeling is the design of scheduling algorithms. This slide shows an
example of an activity scheduling process. Activities are scheduled consistently, such that they
fit into available time windows, taking into account expected travel times. For mandatory
activities in which arrival times are important, departure times may be backed out using the
expected travel time from skims (by the expected travel mode). Higher-priority activities are
typically scheduled before lower priority activities. Residual time windows refers to the time
left over after activities have been scheduled, and provide opportunities for further activity
engagement. Different modeling systems use alternative scheduling processes, but the outcome
(that activities are scheduled consistently and realistically with no overlaps) should be observed
across all systems. No person can be in more than one place at one time.
311
Page 47
HOV Lanes
Light-Rail
Heavy Rail
Commuter Rail
47
We will discuss the treatment of modes in Webinar 10. Different modes imply different model
designs. This depends on how precisely one needs to model access modes and sub-modes. These
range from simpler definitions (like auto, transit) to more precise descriptions such as HighOccupancy vehicle lanes, light-rail transit, bus rapid transit, heavy rail and commuter rail.
Examples of each mode are shown on this slide.
312
Heavy rail: Higher-speed passenger rail cars operating singly or in trains of two or more
cars, on fixed rails in exclusive rights-of-way excluding other vehicular and foot traffic.
Commuter rail is passenger train service operating between a central city, its suburbs,
and/or another central city.
313
Page 48
Treatment of modes
Tour mode versus trip mode
Tour mode is the overall mode for the tour
Trip mode is the actual mode chosen for each trip on the tour
Other considerations
Trip 3: Walk
Zone 3
48
Activity-based models can treat modes simply or can address modes with a great amount of
detail, as described on this slide. A key aspect of activity-based models is the difference between
tour mode and trip mode. The trip mode refers to the mode used for each trip on a tour. This is
observed in household survey data when a respondent reports their travel for the day. Ultimately,
trips are assigned to transport networks based upon the trip mode. The tour mode is an abstract
concept used to classify the mode for the tour as a whole. Typically tour mode is based upon the
modes used for each trip. For example, the diagram shows the same work tour previously shown,
with each trip mode labeled. The tour mode is walk-transit, even though one of the trips (the last
one) is made by walking. The tour mode influences the location of stops on the tour and the
modes of the trips used for the tour.
314
Page 49
Sequencing Models
1.
49
One of the key differences in activity-based model system designs is the use of different
algorithms for sequencing models. This slide shows one way of sequencing tour destination,
mode, and stop location choice. In this example, the primary destination for the tour is selected
first. This selection is based upon the composite utility, or mode choice log-sum, for the tour
mode choice model. Once the primary destination is chosen, the departure time from home, and
arrival time back at home, is chosen. Next, the primary mode is chosen for the tour. Finally, the
locations of stops on the tour are determined, taking into account the additional travel time and
cost required to access the stop based upon the location of the tour origin and primary
destination. We will talk more about model sequencing in Webinar 10, and the pros and cons of
various sequencing options.
315
Page 50
Network Integration
Activity-Based Demand Model
Output: Trip Lists
Feedback
DTAZ
OTAZ
1DTAZ 2
3
4
5
6
7
8
9
10
1OTAZ 1.941 3.402 5.383 6.034 6.545 7.616 8.777 8.098 3.559 5.7310
DTAZ
2 1 3.811.941.893.403.685.384.266.035.256.547.787.618.548.777.878.095.643.557.875.73
OTAZ
1DTAZ 2
3
4
5
6
7
8
9
10
3 2 4.823.813.411.892.123.683.514.265.045.256.697.787.288.545.967.875.265.647.727.87
1OTAZ 1.941 3.402 5.383 6.034 6.545 7.616 8.777 8.098 3.559 5.7310
4 3 6.694.824.023.413.112.122.183.513.475.045.366.695.877.285.465.966.005.268.427.72
2 1 3.811.941.893.403.685.384.266.035.256.547.787.618.548.777.878.095.643.557.875.73
5 4 8.986.695.524.025.193.114.032.182.333.475.595.367.005.877.575.469.376.00
10.62
3 2 4.823.813.411.892.123.683.514.265.045.256.697.787.288.545.967.875.268.42
5.647.727.87
6 5 8.448.987.155.525.865.195.814.034.582.332.555.593.787.006.507.578.859.37
10.31
10.628.42
4 3 6.694.824.023.413.112.122.183.513.475.045.366.695.877.285.465.966.00
5.26 7.72
7 612.408.449.307.157.535.866.885.817.154.585.022.554.243.788.206.50
11.168.85
11.78
10.31
5 4 8.986.695.524.025.193.114.032.182.333.475.595.367.005.877.575.469.37
10.628.42
6.00
8 711.44
7.699.305.357.535.586.886.957.157.355.027.744.244.658.208.91
10.08
11.16
11.78
612.40
10.31
5 8.448.987.155.525.865.195.814.034.582.332.555.593.787.006.507.578.859.37
10.62
9 8 5.53
6.35
6.35
8.98
10.13
10.34
11.07
10.55
4.25
5.90
7.69 5.35 5.58 6.95 7.35 7.74 4.65 8.91
10.08
711.44
11.16
11.78
612.408.449.307.157.535.866.885.817.154.585.022.554.243.788.206.50
8.85
10.31
10 9 7.855.537.846.359.246.35
11.818.98
13.80
12.38
13.14
13.68
6.824.253.285.90
10.13
10.34
11.07
10.55
8 711.44
10.08
12.407.699.305.357.535.586.886.957.157.355.027.744.244.658.208.91
11.16
11.78
10 9 7.855.537.846.359.246.35
11.81 13.80
12.38
13.14
13.68
3.28
10.13
10.34
11.07
10.556.82
8 11.44 7.69 5.358.985.58
6.95
7.35
7.74
4.654.258.915.90
10.08
10 9 7.855.537.846.359.246.35
11.818.98
13.80
12.38
13.14
13.68
10.13
10.34
11.07
10.556.824.253.285.90
10
7.85 7.84 9.24 11.81 13.80 12.38 13.14 13.68 6.82 3.28
Network
Assignment
50
Among the current generation of activity-based models now in use, network integration is fairly
straightforward and not too dissimilar from trip-based models. The activity-based demand model
will create a list of activity-travel events that looks very much like a travel diary!
From these activity-travel records, a trip list is formed, representing vehicle trips for persons
traveling in autos. This trip list includes origins and destinations as well as very specific (minuteby-minute) trip departure times. The trip list may also include contextual or person-attribute
information that might be used for classifying the traveler. For example, in addition to mode, the
trip might be identified by the activity purpose at the destination end and possibly by the drivers
income group, both of which could be used to assign the trip to a value of time grouping. This
is not a feature of the early activity-based modeling systems, but is becoming an increasingly
common design feature of more recent model designs for the purposes of modeling
heterogeneous responses to tolls and fares.
This user class information, along with time of day and origin-destination identifiers, would then
be used to group trips from the trip list into trip tables. Yes, these are the same style of trip tables
316
that should be familiar to all travel demand modelers. One difference, of course, may be that
activity-based models are trending towards more assignment time periods user classes based on
value of time.
Once highway and transit assignments are run, travel time and cost skim matrices are
produced, and fed back to the activity-based model. Current activity-based models use speedfeedback-loop systems very much like trip-based models and use the same network assignment
software. We will cover network integration in greater detail in Webinar 11. It is also worth
mentioning that there are a couple of active research projects going on right now on the
integration of dynamic traffic assignment (DTA) and activity-based models. This includes the
SHRP2 C10 program, the SimTravel project being conducted by Arizona State, University of
Arizona, and U.C. Berkeley, and independent investigations by the San Francisco County
Transportation Authority.
317
Page 51
PID
1
1
2
3
TID
1
2
1
1
PUR
2
1
4
2
MOD
1
2
1
4
SB
0
1
0
1
SA
1
0
0
1
OTAZ
943
943
943
943
DTAZ
987
731
952
565
S1TAZ
0
856
0
698
S2TAZ
964
0
0
982
TLOR
1
3
1
1
TLDS
3
3
2
2
5.0%
Trip Tables
Frequency
4.0%
Observed
Estimated
3.0%
2.0%
1.0%
56
59
50
53
41
44
47
32
35
38
23
26
29
17
20
11
14
0.0%
Peak Highway Travel Time (minutes)
TAZ Accessibilities
transitPeakTotal
Maps, Graphics
<0
0-1
1-2
2-3
3-4
Assignment
Other Summaries
4-5
5-6
7-8
8-9
9 - 10
10 - 11
11 - 12
51
12 - 13
13 - 14
14 - 15
> 15
In addition, the design of the simulation software itself is an important consideration. Since
activity-based models forecast using Monte Carlo simulation, developers have created
mechanism by which the user can fix random number seeds and, in some places, freeze certain
model components to support New Starts analysis or other analyses in which it is desirable to
hold certain model outcomes constant while varying just one, like mode choice.
In Webinar 12, we will cover the software used to implement activity-based models, and the use
of models in forecasting. This slide illustrates how disaggregate data produced by the model,
which includes household, person, tour and trip level information, can be summarized and
analyzed in a variety of ways. Familiarity with the data produced, and the underlying models that
produce that data, is essential for properly interpreting model output and using it to evaluate
policy scenarios. Because activity-based models produce such a large amount of output, it may
be difficult to know what to expect as normal, particularly when looking at various tour-based
measures, activity durations, and other measures that may be new to trip-based modelers.
318
Visualization tools, combined with training, documentation, and experience examining results
are keys to success.
319
Page 52
52
So, lets see what it looks like when we begin to put these different design elements together.
Were going to look at three different model designs, two that are currently in use and one that is
now under development. This slide shows the DaySim model design for Sacramento Association
of Governments (SACOG). Similar models have been implemented for San Francisco County
Transportation Authority (SFCTA) and Denver Council of Governments (DRCOG), with models
under development for Puget Sound Regional Council (PSRC - Seattle MPO), in San Joaquin
and Shasta Counties, California and in Jacksonville and Tampa, Florida.
In this model system long-term and mobility choices appear at the top of the diagram. This
would seem to be a fairly standard treatment of these modeling steps across different platforms.
After this we have activity generation and scheduling steps. This model system is characterized
by an enumeration of possible combinations of tours of different types, which define individual
day patterns. We looked at an example of this earlier. There are literally hundreds of
combinations represented in the model system, which provide a strong conditioning for the
models to follow. Downstream models include models to determine the exact number of tours of
320
each type, destinations, mode and starting and ending times of tours. Given the establishment of
tours and their primary destinations, modes and start and end times, the next model to be applied
determines whether there is room left over the in schedule for intermediate stops on the tour. For
intermediate stops, we then choose locations, modes and departure times.
321
Page 53
Coordinated Travel
Regional ActivityBased modeling
Platform (CT-RAMP)
53
This slide shows the CT-RAMP model diagram for San Diego Association of Governments
(SANDAG). Similar models have been developed for Mid-Ohio Regional Planning Commission
(MORPC), Tahoe Regional Planning Agency (TRPC), Atlanta Regional Commission (ARC),
Metropolitan Transportation Commission (MTC), and under development for Maricopa
Association of Governments (MAG), the Miami region, and Jerusalem, Israel.
In this model system, we also see population synthesis and long-term and mobility choices at the
top of the structure. This model system is characterized by the identification of mandatory
activities (work and school), which are prioritized in schedule creation. Persons are assigned
daily pattern types, which are defined rather simply as being mandatory, non-mandatory or
home. It is also characterized by explicit modeling of joint activities, through a series of models
that determine joint tour frequency, party size, and individual participation. Time windows play
an important role in this model system as well, with mandatory activities scheduled first,
followed by joint activities, and finally independent discretionary activities. Frequencies, time of
day, and destination decisions are made at this level. This is followed by tour modes, stop,
322
frequency, stop location and departure times. Trip modes, auto parking and network assignment
are the final steps.
323
Page 54
54
This slide shows the Comprehensive Econometric Micro-simulator for Daily Activity-travel
Patterns, by Dr. Chandra Bhat of University of Austin-Texas. The model has been applied to the
Dallas/Ft-Worth region, and a version of the model is currently under development for the
Southern California Association of Governments (SCAG).
This model system is characterized by an activity generation and allocation step the
predetermines the amount of activity participation, expressed in terms of time spent in an
activity. This is done through a discrete-continuous model formulation. Household-level
responsibilities for ridesharing are also determined at this stage. The system then proceeds
through a series of scheduling steps, which prioritize certain types of tours or partial tours. For
example, round trip work-commute characteristics are identified first, followed by any drop-off
and, then, pick up tours for non-workers who are escorting children to school. CEMDAPs
scheduler then proceeds through a number of other tour types, ending with independent
discretionary tours for children. Although not shown here, within each tour scheduling group,
there are the usual activity stop generation, location, mode and timing decisions. In addition,
324
population synthesis and long-term and mobility choices are determined through a separate
module, not shown here.
325
Page 55
55
The practical implications of different model system designs come down to a tradeoff between
sensitivity to certain policy inputs versus the cost of running and maintaining the modeling
system. By cost, were talking about both time and money. The more complexity that is built into
a modeling system, the more it costs to develop, the more input data needs to be developed and
maintained, and the greater the computational load. More complex models will take longer to run
and, accordingly, will require investments in hardware and software to make it feasible to run
them in a production context. In addition, a more complex model could be more difficult to
interpret, although, this can be overcome by careful design of output reporting capabilities and
features, and of course user training and documentation.
The developers of these modeling systems are well aware of these tradeoffs. They have carefully
thought through their model designs, choosing to add more detail where they felt it was most
important, and leaving out certain details where they felt it was less important. These design
decisions tend to be fairly nuanced. For example, one decision is whether to model at a fine level
of spatial resolution, such as parcels, or to use a more aggregate measure of spatial units such as
326
TAZs. The use of parcels has the potential to provide the best estimates of non-motorized travel
times, but adds complexity to network loading, calculations of travel times and impedances, and
use of skims.
Another example might be whether to model intra-household interactions explicitlyjoint
activity participation and pick up and drop off eventsor to assume that these events happen,
but that it is sufficient to treat each persons experience independently. Modeling intra-household
interactions explicitly is more realistic, but adds a layer of model components due to the need to
generate and coordinate patterns between individuals. In both of these examples, the question is
whether the added complexity produces enough of a meaningful difference in model results to
merit the extra costs. It is very important for agencies to on their modeling needs and priorities
when developing a new model system.
327
Page 56
Ongoing Research
Model structure
56
Practical activity-based travel modeling is relatively new, and there are a number of areas of
ongoing research. Activity-based model structures are being extended in novel ways, to
explicitly address in-home versus out-of-home activities, full treatment of joint travel, household
car allocation and fleet models, integration with dynamic traffic assignment (DTA), land-use
models, and emissions models.
There is also a very active body of research in the realm of discrete choice models. Model
structures are being explored that address more, inter-related alternatives. Models have been
developed that simultaneously model both discrete and continuous alternatives, such as number
of vehicles owned and amount of miles driven on each. There are many models that address the
issue of user heterogeneity, such as cost sensitivities that are expressed as a distribution of cost
parameters rather than as an average parameter that applies to everyone. Finally, some models
explicitly model user preferences for packages of alternatives or attributes of alternatives, rather
than relying solely on socio-economic variables to explain different sensitivities. We will explore
some of these topics in subsequent webinars.
328
329
Page 57
57
Discuss how household activity-travel diary data is used to define activities, tours, and
daily patterns
Describe how choice model structures are used to represent key aspects of activity-based
model generation and scheduling
Describe how discrete choice models are used and applied in activity-based modeling
systems
Discuss the various design decisions that are important to the development of activitybased modeling systems
330
Page 58
331
Page 59
February 2
February 23
March 15
April 5
April 26
May 17
June 7
June 28
July 19
August 9
August 30
September 20
59
332
Page 60
60
333
locations in the baseline alternative might not make sense given the location of service in the
build alternative. Different model platforms have different techniques for addressing such issues.
Another case where you might hold some choices constant is when you want to know short-term
responses to policies. You might hold long term choices like auto ownership and work location
constant and only re-simulate mode and route choices to evaluate short-term impacts.
Is the Poisson probability distribution used in activity-based models?
John: The Poisson distribution is usually used to model a random arrival process. It is
occasionally used to model trip generation processes, as a counting process. It isnt very
commonly used. Not many of the other choices in an activity-based model are amenable to a
counting process.
Is the mode for a tour determined comparing utilities to the primary destination only, or is travel
to intermediate stops factored into the choice of tour mode? And are trips between intermediate
stops constrained depending on the choice of tour mode?
Joel: Yes, we are typically choosing a tour mode based on round trip travel times and costs to the
primary destination, though we can use information from the stop location choice model to
influence the tour mode. For example the accessibility to stops (the stop location choice log-sum)
can influence the number of intermediate stops, and this information can also be fed back to
influence the tour mode. It is typically found that the mode for the tour is more influenced by
round trip times and costs than characteristics of the traveler or the tour. The second part of the
question is about mode switching. Most models do allow for mode switching. The tour mode is
an abstract concept. The trip mode is what is actually observed in the activity diary. A lot of
mode switching occurs in surveys, for example switching between shared ride and drive alone
auto modes on the tour, so we do need to allow for that switching to happen.
335
336
Page 1
Activity-Based Modeling
Session 5: Population Synthesis and Household
Evolution
337
Page 2
Acknowledgments
This presentation was prepared through the collaborative efforts of Resource
Systems Group, Inc. and Parsons Brinckerhoff.
Presenters
John Gliebe, Peter Vovsha
Moderator
Stephen Lawe
Media Production
Brian Grady
Resource Systems Group and Parsons Brinckerhoff have developed these webinars
collaboratively, and we will be presenting each webinar together. Here is a list of the persons
involved in producing todays session.
John Gliebe and Peter Vovsha are co-presenters. They were also primarily responsible for
preparing the material presented in this session.
Stephen Lawe is the session moderator.
Content development was also provided by Bhargava Sana, Joel Freedman, and Maren
Outwater. John Bowman and Mark Bradley provided a review of the material.
Brian Grady was responsible for media production, including setting up and managing
the webinar presentation.
338
Page 3
February 2
February 23
March 15
April 5
April 26
May 17
June 7
June 28
July 19
August 9
August 30
September 20
3
For your reference, here is a list of all of the webinars topics and dates that have been planned.
As you can see, we are presenting a different webinar every three weeks. Three weeks ago, we
covered the fourth topic in the seriesActivity-Based Model Frameworks and Techniques. This
session provided some of the concepts fundamental to activity-based modeling, including how
activities and tours are represented in data; and how various choices structures may be used to
model particular dimensions of activities, tours, and travel patterns. We also covered important
aspects of discrete choice modeling and model implementation using Monte Carlo simulation
methods. In the second half of that webinar, we discussed activity-based model design and
highlighted some of the key trade-offs that model developers consider. Todays session is the
second of nine technical webinars, in which we will cover population synthesis and household
evolution models.
339
Page 4
Learning Outcomes
By the end of this session, you will be able to:
Describe a synthetic population
Describe the methods used to synthesize a population
Describe the process of household evolution
In todays session, we will be covering the basics of activity-based modeling. At the end of this
session you should be able to:
340
Page 5
Session Outline
In this webinar, we will describe a synthetic population and how its place in the overall structure
of an activity-based model system. This will be followed by a discussion of considerations for
specifying a synthetic population generator. Next, we will cover the methods used to create a
synthetic population. Well then look at some examples of synthetic population generators that
have been implemented. Finally, the last part of todays webinar will introduce a class of
emerging methods in population synthesis known as household evolution models.
341
Page 6
Terminology
Socio-demographic attribute
Controlled attributes
Disaggregate household sample
Seed data
Iterative Proportional Fitting (IPF)
Drawing a sample
Uncontrolled attributes
Household evolution
In todays webinar, we will discuss some of the essential methods in population synthesis. First,
wed like to define some of the terms that we will be using frequently throughout the
presentation.
342
343
Page 7
Individual Choices
First, it might be instructive to review how populations are represented in trip-based models and
contrast them with disaggregate activity-based models
In trip-based models, we are dealing with aggregate households grouped in transportation
analysis zones (TAZ). We apply trip production rates to groups of households, and trips are
generated and aggregated by TAZ.
In activity-based models, we maintain households in disaggregate form. Individual choices are
simulated, and individual activities, trips, tours are generated and scheduled.
344
Page 8
Synthetic populations are essential to simulating individual activity-travel patterns. The design of
the synthetic population should support the design of the activity-based model and provide the
variables it needs. In addition, the activity-based model should only rely on information that can
be realistically provided in the synthetic population.
345
Page 9
Land Use/Spatial
Database
Transportation
Networks
Population
Synthesis
Individual
Household Agents
Population
Aggregation
Accessibility
Measures
Individual Person
Agents
Activity Pattern &
Tour Generation
Activity & Tour
Scheduling
Activity-Based Modeling: Activity-Based Model Basics
In an activity- based model, population synthesis is a modeling step, much like activity pattern
and tour generation and activity and tour scheduling steps. Census-related data is combined with
geographic data to create synthetic households, with specific locations throughout the region of
interest.
Population synthesis is the first step in the modeling process and produces individual household
agents and individual person agents that are the subjects of the simulation. These household and
person agents make decisions that through the activity generation and scheduling process.
Since synthetic households/persons are situated spatially, they may be combined and aggregated
at TAZ level or some other spatial units. Some of the accessibility measures used in activity
pattern and tour generation and scheduling models are derived from aggregations of households
within a buffer region, usually specified to consider transportation network distances, travel
times and costs. These accessibility measures, which we will discuss at length in the next
346
webinar, also appear as explanatory variables in activity pattern and tour generation and
scheduling models.
347
Page 10
Population
Long-Term
Mobility
Activities
Person and household attributes
should be defined to reflect
mode availability and
propensities.
Modes
Space
Tour Patterns
Time
10
Lets consider how decisions regarding the synthesis of a population for the purposes of
simulation might affect other aspects of model system design.
This diagram serves as a backdrop for describing the relationships between key design elements
in activity-based modeling. These elements include: defining the population, modeling longterm and mobility-related choices, defining activity types, defining modes, defining tour patterns
and an entire day-pattern elements, as well as the treatment of space and accessibility and
treatment of time. We discussed each of these design elements in the previous webinar on
activity-based modeling frameworks and techniques, and we will devote an entire session to each
one of these elements starting today (population), and over the next six webinars.
First, I mentioned earlier that the method of population synthesis usually provides only the basic
characteristics of the population. It is usually desirable to model other attributes of households
and persons as long-term or mobility decisions. This enables us to reflect policy-sensitivity into
these choices. For example, household auto ownership is related to income, number of drivers,
348
and presence of workers. It may also be related to some notion of accessibility, which we will
talk about in later webinars, that is derived from the level of service provided by the
transportation network. In addition, we typically want to model certain person attributes as longterm decisions, such as workplace, school and college locations. Others may be thought of as
mobility decisions, such as whether to participate in a TDM, buy a transit pass, buy a
transponder, or whether an individual will benefit by employer-subsidized parking.
Second, complementary definitions of population attributes and activity types and tour types are
important. For example, person attributes, such as whether a person is a worker or student,
determine the availability of work and school activities and home-based tours for these purposes.
In addition, household composition variables and how they are represented in the synthetic
population (presence and ages of children) determine the ability to model household drop-off and
pick-up events, carpooling. They also affect the propensity and structure of joint activity types
and participation, and the formation of joint travel tours.
Finally, population attributes impact mode availability. It is important to identify persons of
driving age. Likewise, household income attributes may be used to segment persons by their
value of time or willingness to pay for travel time savings.
349
Page 11
Person Type
Age
Work Status
School Status
Full-time worker
18+
Full-time
None
Part-time worker
18+
Part-time
None
Non-working adult
18 64
Unemployed
None
Non-working senior
65+
Unemployed
None
College student
18+
Any
College +
16-17
Any
Pre-college
Non-driving student
6 16
None
Pre-college
Pre-school
0-5
None
None
Model segmentation
Summarize outputs
Explanatory variables in models
Constraints on available alternatives
11
We showed this slide in the last webinar on activity modeling frameworks and techniques, but it
is worth reviewing again here because it provides a good summary of the role that person and
household attributes play in the overall model design. Although we are modeling persons in
disaggregate form, it is often useful to create person-type categories. This table shows an
example of person-type categories and definitions. Person-type categories can be used for a
number of purposes:
As a basic segmentation for certain models, such as daily activity pattern models.
To summarize and compare observed versus estimated data and calibrate models.
As explanatory variables in models.
As constraints on alternatives that are available; for example, work and school activities
are only available to workers and student; and driving is restricted by age.
Although continuous values of age and income are available, properly grouped categorical
variables often result in a better model fit and efficiency. Determining the proper cutoff points
350
for categorical variables related to age and income are usually derived in the model development
process empirically through descriptive statistical work. For example, a community with a large
population of retirees might use different age groupings than a college town. Not having proper
cutoff points could result in important market segments being under-represented in model
specifications and estimated models not explaining as much behavioral variation in the
population as they could.
Lets look at a realistic example of a policy study and the specification of population attributes,
such as income groups, might affect our analysis.
351
Page 12
Build Alternative(s)
12
Lets consider a transportation planning and policy project that might be faced by an MPO or
DOT and how population synthesis fits into the picture. We may revisit this example in each of
the next several sessions to come and discuss how the topic of that session relates to this
particular example.
For this scenario analysis, we will be considering a number of alternatives: a no-build alternative
and a various configurations of the build alternative. In the no-build alternative the bridge has 4
lanes (2 in each direction), there are no tolls, and the transit fare stays the same all day. In the
various build alternatives, there are 6 lanes on the bridge. In some alternatives the two additional
lanes will be HOV lanes all day, while in other alternatives the two additional lanes will be HOV
lanes only during peak periods. In addition, in some build alternatives there will be a new toll
that is the same across the entire day, while in other build alternatives there will be a toll that will
be only applied during peak periods, or when certain levels of congestion occur. Finally, in the
build alternatives regional transit fares will be higher during peak periods.
352
353
Page 13
13
For this bridge example, the synthetic population plays a major role in determining how demand
will vary through the population. The set of methods that we will focus on in this webinar,
generally known as population synthesis provide some of the answers. Of particular interest to
this example, population synthesis provides direct representation of the income levels of
individual travelers and household structure. Thus, it is important that the synthetic population
covers the entire income range in the real population and, that those income levels are properly
correlated with other important attributes, such as age of householders, household size, number
of workers, and auto ownership. It is also important that households with these attributes be
distributed properly in space. For example, we need to make sure that low-income households
and zero-car households are placed in neighborhoods that match their occurrence in the real
world.
354
Page 14
HOV
Potential?
Household records
TAZ
1433
1937
77
18
HHID
16670
17392
232
5042
Age of
Household Number of
Head
Persons
1
2
1
2
1
3
1
3
Income
Group
1
2
3
2
Attribute values to
be updated by
Long-Term models
Person records
Household
TAZ
ID
Person ID
77
232
1
77
232
2
77
232
3
Presence
of Children
1
1
1
1
Works
From
Home
Age
22
24
1
Employment
Status
1
1
0
1
1
0
Autos
Owned
0
0
12
3
Number of
Workers
1
1
2
2
Hours
Worked
Gender Is Student per Week
2
0
9
1
0
45
2
0
0
14
What does a synthetic population look like? This slide shows an example of output from a
population synthesizer. Prior to their use in the simulation, synthetic populations are represented
in data tables, often in a relational database or some equivalently structured file system.
Typically there are separate tables for households and person records. The format of the data
tables may be familiar to those who have worked with household travel surveys. Person records
are linked to household records through ID numbers. Attributes of synthetic households and
persons may be found in each record. The TAZ ID identifies the geographic location of this
synthetic household. Putting the right households in the right locations is critical to model
performance. The values of the other attributes are those that are typically available from the
Decennial Census or American Community Survey (ACS) for control variables. For other
variables, usually uncontrolled, the attribute values in the disaggregate sample are what is
available. This disaggregate sample often comes from PUMS data, but could also be from a
household survey. Well talk more about data sources in a few minutes.
355
Some of these variables and values may need to be updated or modified in some way. For
example, while it might be desirable to have household auto ownership generated by the
population synthesis, and this is useful as a calibration check, we typically prefer to model
household auto ownership or availability through a separate process. This is so that it can be
made sensitive to transportation system accessibility and other important policy variables that
might change in future scenarios. We will cover auto ownership along with other long-term
household and medium-term mobility decision models in Webinar 7, two sessions from now. For
our bridge expansion scenario, in addition to household auto ownership, we may also be
interested in modeling whether an individual has a transit pass or a transponder, both of which
are typically modeled as mobility decisions.
Similarly, we tend to model an individuals decision to work from home as a daily activity
pattern alternative that might be influenced by not only demographic factors, but also
transportation system and other policy variables. Activity pattern generation is the subject of
Webinar 8, where you will see how important individual household attributes become in the
prediction of daily tour patterns. As stated a few minutes ago, the bridge analysis will be greatly
affected to by the number of commuters, and the potential demand for HOV usage. So, knowing
whether someone is likely to telecommute or whether they live in a household with other
workers, who might share rides, would be key inputs to that analysis.
Other variables may need to be augmented or transformed for our use. Here, we have used
PUMS variables that indicate the number of hours a person works per week, which we might
later transform to a variable that indicates whether a person works full-time or part-time. In this
example, we control for household income by group and that is what shown, but we might also
want to draw from our sample a continuous value of income. This might help us be a bit more
precise in modeling incomes for the value-of-time calculations that are so important to the bridge
study.
Finally, since some activity-based models are designed to utilize a level of geography that is
finer in resolution than the spatial units used for control variables, there may need to be a
separate sub-allocation process to place households on parcels.
356
Page 15
15
There are three major steps in creating a synthetic population, the first of which is specifying the
inputs to the processthe control variables and sample households as well as the level of
geographic resolution. Specifying the control variables is essential. In addition, there is often an
additional step of specifying additional, uncontrolled variables to be added to the synthetic
population.
The second major step is actually running a program that produces the synthetic households.
Synthetic population generators may be packaged with an activity-based travel model. In
addition, some standalone population generators available, and we will talk about these a bit
later.
The third major step would be transforming the model-generated outputs into characteristics of
the population that will be used throughout the rest of the model system. This would be doing
some of the things we just discussed, such as creating categorical variables out of continuous
357
variables, reformulating income, or allocating households from the zonal level to a finer level of
geographic resolution, such as a parcel.
Next, we will talk about each one of these steps in more detail. In order for you to better
understand the first step shown here (specifying a target distribution), you need to know what is
coming downstreamhow the synthetic population generator works.
358
Page 16
HH by Income
HH by Size
SF AGE_HH
0
24
1
23
1
43
0
32
1
34
0
49
1
67
1
15
0
12
Household File
HID SIZ INC WRK
1
1
1
1
3
1
1
0
5
1
1
1
SF
0
1
1
AGE_HH
24
43
34
Person File
HID PID AUT INC WRK GEN AGE EMP
1
1
1
1
1
0
24 1
3
1
3
1
0
0
43 1
5
1
0
1
1
1
34 0
16
So, lets briefly review how a population synthesizer works. We showed this slide during
Webinar 4. It provides a good conceptual view of a general process.
In the first step, marginal control totals for each zone are created along with a sample of the
correlation between attributes, usually from a disaggregate sample. It is common to refer to these
control totals as marginal control totals because they represent the margins of the joint
distribution. They are represented here as row and column target values. It is also common to
refer to this sample data as seed data or the seed matrix if structured as a matrix, as in this
example.
In the second step, the aggregate characteristics of the synthetic population are defined, in the
form of a joint multi-way distribution of household attributes for each zone.
In the third step, individual person and household records are drawn from a sample of
households, such that when we sum them up, they meet the joint distribution defined for the
zone. So, the construction of joint multi-way distribution of households provides a target
359
distribution. Within each cell in the matrix are the number of sample households that match those
control attributes that we need to draw from the disaggregate sample.
360
Page 17
17
We will discuss more about the technical details of population synthesizers in few moments.
First, we need to return to the top of this flow chart and cover the essential steps involved in
creating the inputs to the population generatorspecifying the target distribution through the
selection of control variables. We will also talk a little about the use of uncontrolled variables.
361
Page 18
Identify which
attributes to control
and at what level
Determine level of
spatial resolution
Household size: 1, 2, 3, 4+ or .. 5+ ?
Income level: Lo/Med/Hi? Quartiles? CPI Deflated?
Presence of children: no/yes by age group?
Age by group: 0-4, 5-15, 16-17, 18-24, 25-64, 65+?
18
The first step in population synthesis is to determine which attributes are most important to the
model design and then to specify those as the variables to be controlled. This includes both
important household and person attributes and should be policy-driven; however, over the years
a fairly standard set of variables has emerged, namely household size, age of householder,
number of workers, and income. The presence of children and person-attributes for age and
gender are also desirable and fairly common.
The next step is to obtain the data needed to summarize the distribution of these attributes in the
population. This might includes Census, American Community Survey (ACS), CTPP, or other
TAZ-level control totals.
It is important to determine how to group the data, so that the resulting population covers the
sample space. What is shown here are just examples. For instance, we might specify age group
control values for person attributes, so that we make sure we draw samples that include preschool age children (0-4), school-age non-driving children (5-15), driving-age high-school
362
students (16-17), college-age adults (18-24), and senior citizens (65+). For some regions, other
groupings might make more sense. In addition, specifying income levels can be tricky, because
of the need to ensure representation from low and high-income households and to adjust for
inflation. In our bridge example and other studies involving equity analysis, income specification
is important to estimating not only willingness to pay for travel time savings, but also for
identifying how different income groups benefit from a proposed project or policy. Since the
model will be used for multiple purposes, it is important not to be too narrow in focus to avoid
specifying variables that will limit other analyses.
Typically, the smallest level of spatial resolution that can be feasibly and reliably used to control
attributes is used. This is usually a function of available data (e.g., Census blocks, block groups,
TAZs). In recent application, control totals are grouped at the TAZ level, which makes things
easier for forecasting. In addition to the data itself, it is typically necessary to use GIS layers
representing blocks, block groups, TAZs and potentially other spatial units to develop
equivalency tables.
363
Page 19
Base year
Decennial Census: ~100% sample
American Community Survey (ACS) summary files:
3% sample, rolling 5-year sample, yields an estimate of ~15% of pop.
Other zonal data developed locally (TAZs)
Census Transportation Planning Products (CTPP)
TAZ-Census Block/Group geographic equivalencies
Forecast year
Regional socio-economic forecasts, growth factor models, allocation
processes
Output from a land-use model
19
A bit more needs to be said about data sources. Here again weve listed some of the typical
sources used in the U.S.
They fall into two general categories based on how they will be used in the population generator:
Sources of control data, which include full-population estimates/counts/forecasts at the desired
level of geographic precision
Sources of disaggregate household/person sample data, which provide a basis for estimating the
correlation between attributes
First, we will talk about data for specifying the target distribution, the control data. Base year
data might include: the decennial Census (2010) is now available, which provide a 100% sample
of the population and is available in summary form as fine as the block level for certain attributes
of interest, such as household size, age of householder, presence of children; and person
364
attributes, such as age and gender. The smaller the geographic units, the closer the synthetic
population should be to the true population.
Other attributes, such as households by income group and number of workers may only be
available in summary form from the American Community Survey (ACS), which has become an
annual 3% sample, that is available in increments of 1, 3 and 5 years.
In addition, some MPOs may develop their own socioeconomic data locally, using some
combination of sources. For example, this could include households living in single- and multifamily housing unit types, as well as seasonal households.
The Census Transportation Planning Products (CTPP) can be used and is under preparation for
2010. It typically includes ACS level data and TAZ-Census geographic equivalencies.
TAZ-Census geographic equivalencies may be developed through GIS overlay analysis if they
have not been created through CTPP or some other source.
Forecast year control data may be derived in any number of ways. The Census is not available
today for future years, so official regional socio-economic forecasts may be used or simple
growth factor models, in combination with some type of allocation process to produce TAZ-level
forecasts. In some regions, a land use model may be used to generate these forecasts.
365
Page 20
20
In most cases, the primary source of disaggregate sample data will be PUMS data, which is now
part of the ACS, and follows the same sampling framework, but provides disaggregate records
for households and persons across numerous different attributes. PUMS is sampled and grouped
according to geographic units, better known as PUMAs. PUMAs cover contiguous areas of
roughly 100,000 population, including persons living in group quarters. For example, a metro
area of 850,000 might be covered by 8 or more likely 9 PUMAs.
In some cases, it might be desirable to use a representative regional household survey for the
disaggregate sample. This is generally not done, because of concerns over the insufficient
quantities and potential sampling bias. It may make sense, however, for generating populations
of certain sub-groups. For example, the NHTS add-on survey for Florida, included seasonal
households, the characteristics of which could be used to generate a synthetic population of these
part-year residents who do not show up in the in the Census of residents.
366
In general, ACS-PUMS provides good representative coverage of most regions and is rigorously
tested and monitored, so it is generally preferred over household surveys for developing
characteristics of the population.
For forecasting, there are particular challenges associated with areas that are likely to change
their land use composition significantly. This may require special analysis to predict how such
areas are likely to change in terms of the types of households expected to live there. Again, this
is where a land use model may be helpful.
367
Page 21
21
Since the target distribution of households is so important, it is worth discussing some of the
nuances of specifying control attributes.
Control attributes may be single or multi-dimensional. At right, weve depicted an example of a
3-way distribution of household size (4 levels), workers (3 levels), and income (4 levels). Multidimensional attributes can be treated as single dimensional attributes with number of categories
equal to the product of the numbers of categories in individual attributes. So, in this example,
which was developed for the SHRP2 C10A Jacksonville model, there would be 48 categories if
all cells were feasible. However, certain combinations are definitely infeasible, because we
cannot have more workers than household members, there only 44 feasible combinations.
The primary advantage is more precise regional control over the correlation between attributes.
The disadvantages are that some of the combinations of attribute levels are rare and may be nonexistent within certain sub-geographies. These sparse combinations may make difficult to find
samples to meet the cell quota during the drawing process. In addition, if a certain cell
368
combination is represented by only one or two households in the sample, those households may
be over-represented in the final population.
369
Page 22
22
It is not possible to control many attribute levels well simultaneously. There are tradeoffs to be
made that can lead to a more efficient design. In this sense, efficiency means explaining as much
variation as possible with the fewest number of variables. The best choices of variables, will be
meaningful attributes that are somewhat orthogonal to each other, which means that their
variance in the population is largely independent.
Conversely, if you have two attributes that are highly correlated, then controlling for both may
not achieve much more than controlling for just one. For example, if in a particular region certain
income categories are highly correlated with the race of the household head, then it may not be
efficient to control for both household income and race at the same time. In our bridge tolling
and transit pricing example, this is something wed have to look at more closely, particularly if
were interested in equity analysis.
370
Page 23
23
The number of control attributes is also important. On one hand, if there are too few control
attributes, the synthetic population may not accurately reflect the true population. In our bridge
example, wed like to control for as many income levels as possible, as well as number of
workers, and race in order to be able to better address concerns of driver willingness to pay,
estimating HOV and transit demand, and assessing equity impacts.
On the other hand, too many control attributes leads to sparse cells in seed data, which may
distort the true joint distribution, or even make it impossible to compute using certain methods.
And, as mentioned earlier, this sparseness may make it difficult to find suitable sample
households in the disaggregate data, or perhaps just one or two sample households, which get
replicated more than they should.
Control attributes also may use different geographic units. So, if control attribute totals are not
accurate at a particular spatial unit, they could be specified at a lower resolution. Of course, it is
best if spatial units nest (Census blocks & block-groups).
371
Finally, different sets of control attributes may be used for base and forecast years, if limited by
forecasting accuracy. This is not necessarily desirable, though. The ability to forecast marginal
control totals should be a consideration when specifying control attributes for this base year.
372
Page 24
Uncontrolled Attributes
Uncontrolled attributes are directly transferred to the
synthetic population when the sample is drawn
All the attributes present in the sample may not be controlled
for since it could affect the match negatively
Such variables are often needed by activity-based models in
addition to the control variables
Should be well-correlated with controlled variables
Examples
Person worker status, student status, race, occupation
24
Although not every desired variable is available at the population level, it is possible to add them
to the synthetic population by drawing them from samples, such as PUMS or a household
survey. To the extent that uncontrolled attributes are correlated with controlled attributes, this
should be a realistic estimate of their presence in the population. In addition, one may also want
to select additional uncontrolled attributes to support analysis needs, such as work and student
status for individuals, race and occupation.
It is important to identify the risk of biased results if uncontrolled attributes that are not
correlated with controlled attributes are used as explanatory variables in the models, or used to
aggregate model results. A real-world example of this is Atlanta Regional Commission, where
race was not controlled, and a backcast revealed that it was not well synthesized. If race had been
used in forecasting to aggregate model results, they would have been biased.
So for our bridge example, we might want to consider controlling for race, and not leaving it up
to the luck of the draw so to speak. Wed also want to conduct a back-casting exercise to make
373
sure our projections for key sub-groupslow-income households, minority households, and
potentially other groupswind up in the right geographic locations.
374
Page 25
Disaggregate Sample
(PUMS, HH Survey)
Fitting/Balancing
Fractional HH Weights
Discretization
Integer HH Weights
Drawing
Synthetic Population
Activity-Based Modeling: Activity-Based Model Basics
25
This diagram summarizes the population synthesis process. The first step is balancing to create a
population distribution, using the disaggregate sample and the control variables as inputs, as we
just discussed.
Here, the algorithm fits the disaggregate sample of households to aggregate constraints (control
totals) at prescribed level of spatial unit control. The most commonly used method for the fitting,
or matrix balancing, is the Iterative Proportional Fitting (IPF) procedure, which we will illustrate
with an example in a few minutes. There are other balancing procedures as well, including list
balancing. The outcome of the balancing step is a set of fractional household weights that
describe the multidimensional distribution of households in the region, a defined by these control
variables.
The next step is to draw a sample to create individual household and person records to match the
balanced distribution. Before we can do this, we need to first convert fractional values in the
resulting multidimensional distributions to integers. This is easy enough to do with a bucket
rounding procedure.
375
After this, we can select individual households from the sample to match the fitted, integerized
distribution. Any specified uncontrolled attributes will also be included in the population
attributes due to their presence in household and person sample records.
376
Page 26
26
The iterative proportional fitting (IPF) method, or some variant of it, lies at the heart of most
population synthesizers. It is used to produce an estimate of the joint distribution of attributes at
the specified levels and spatial units. It is also a bit easier to visualize than the list balancing
method and employs the same principles of adjusting the sample to meet target distributions, so
we will use it as an example to illustrate this process.
The basic steps of the IPF procedure are as follows:
377
Page 27
Low
Adjustment
3 or more
Total
Household
Size Control
20
Total
Household
Income Control
40 Controls 60
Household Size
Sample
50
Controls
30
27
Here is an example of an IPF procedure that uses a matrix approach, with only two control
variableshousehold size (3 levels) and household income (2 levels). It is an unrealistically
simple example.
The disaggregate sample is tabulated and summed to fill in the number of observations that
correspond to each cell within the 3-by-2 matrix.
The marginal controls totals are represented row and column target values for household size and
income, respectively. As you can see, the control totals for each dimension are significantly
larger than the initial sums from the sample data.
IPF works when the joint attributes are defined as cells, the universe of the joint distribution
being the entire set of cells, and each control total is associated with some subset of the cells. An
iteration consists of adjusting for all controls, one-by-one in sequence. In the extreme case, each
household in a representative sample could serve as a cell, which is essentially how PopGen
works.
378
379
Page 28
Low
Adjustment
20/5 = 4
50/5=10
3 or more
30/4=7.5
Total
Household
Size Control
20
50
30
Total
Household
Income Control
40
60
Household Size
28
To begin the procedure, we use the ratio of the control total divided by the summed total for each
row to create row factors.
380
Page 29
Household Size
High
Low
Total
Household
Size Control
Adjustment
2x4=8
12
20
20
10
4x10=40
10
50
50
3 or more
7.5
1x7.5=7.5
22.5
30
30
Total
Household
Income Control
40
60
29
Applying these row factors, we get new cell values and new row sums that exactly match row
control totals for household size.
381
Page 30
Adjustment 40/55.5=0.7
2
Low
Total
Household
Size Control
60/44.5=1.3
5
12
20
20
40
10
50
50
3 or more
7.5
22.5
30
30
Total
55.5
44.5
Household
Income Control
40
60
30
The row adjustments also changed the column totals for the two income levels, but they still do
not match their control totals. So, we take the ratio of control total to column sum to create
adjustment factors.
382
Page 31
Low
0.72
1.35
8x0.72=5.8
Total
Household
Size Control
12x1.35=16.2
22
20
28.8
13.5
42.3
50
3 or more
5.4
30.3
35.7
30
Total
40
60
Household
Income Control
40
60
Household Size
Adjustment
31
Applying these column factors to the cells in the matrix, we obtain column sums that match
income target values exactly. However, we can see that our row sums are now a bit off from the
household size target values for each row.
383
Page 32
Low
Total
Household
Size Control
4.51
15.49
20
20
31.70
18.30
50
50
3 or more
3.79
26.21
30
30
Total
40
60
Household
Income Control
40
60
Household Size
Adjustment
32
This iterative row and column factoring continues until, at Iteration 13, we have a final set of
values in the 3-by-2 table that produce row and column sums that match BOTH household
income and size control totals, respectively. Although this example showed a 2-dimensional
table, this same process may be extended to 3 dimensions and greater. It is guaranteed to
converge, provided that there are not too many cells in the table with values of 0.
384
Page 33
33
Once weve got the multi-dimensional target distribution figured out, the next step is to draw
households from our sample to match that distribution. The general steps in this process begin
with calculating selection probabilities for each household in the sample, based on the attributes
of the household and the number of such households in the joint distribution.
Next, we draw households from a sample, based on the selection probabilities to match target
numbers by each household type in the joint distribution. The most common method for doing
this is a simple random drawing based on the probabilities weights, using Monte Carlo methods.
(For those of you unfamiliar with Monte Carlo selection, we discussed this in Webinar 4.)
Other, potentially more efficient methods involve cycling methods in which we make sure we
sample from all of the available households that meet the joint distribution criteria, taking each
one in turn, rather than leaving it to chance.
Once we have drawn households to fill meet our target population distribution, we output a
synthetic population consisting of all persons belonging to the households drawn.
385
386
Page 34
List balancing
Involves applying IPF on individual households in a list
More complex algorithm
Both household and person attributes can be simultaneously
controlled
Reduces sparse matrix problems
Activity-Based Modeling: Activity-Based Model Basics
34
List balancing involves applying IPF to individual households in a list. List balancing is a more
complicated to program, but has two advantages of over matrix balancing. First, both household
and person attributes can be simultaneously controlled. Second, it is easier to eliminate matrix
cells that are irrelevant or combine those that have extremely low incidence rates.
387
Page 35
35
Two emerging methods in drawing a sample are intelligent draws and discretizing. Intelligent
drawing eliminates the need to develop a target multi-way distribution through a separate fitting
and balancing step, so no IPF process. Instead, households are drawn directly from the
disaggregate sample to match control totals. Drawing is informed by the extent to which a
sample household would contribute to the match of both household and person controlled
attributes. The challenge is in developing selection criteria and algorithm, which is why this is
considered to be an emerging method, although there have been at least two implementations of
this: ARC population synthesizer and FSUTMS in Florida.
388
Page 36
36
The other method is referred to as discretizing, which may be thought of as intelligent bucket
rounding. In this method, households are assigned weights, which may be developed through
IPF or an entropy maximization method. The weights should be proportional to their probability
of selection, if Monte Carlo draws were used. However, Monte Carlo draws are not used. Rather
the fractional weights are integerized using methods similar to bucket rounding, which seeks to
preserve totals. Each integer instance of a household can then be used directly in the synthetic
population. This method is similar to list balancing without the drawing step. One example of
this has been implemented in San Diego. This approach has the advantage of ensuring full use of
the sample space when selecting households, which is not guaranteed in Monte Carlo sampling.
389
Page 37
390
Page 38
List balancing
Fitting only
PopSynWin
(U. Illinois-Chicago)
CEMSELTS
(U. Texas-Austin)
TRANSIMS
(FHWA)
ALBATROSS
(TU-Eindhoven)
MORPC
(PB)
PopGen (Arizona St.) SANDAG PopSyn
ILUTE (U. Toronto)
(PB)
Drawing only
ARC PopSyn (PB)
FSUTMS
(U. Florida)
38
This table is a list of known population synthesizers, or the modeling packages of which they are
a part. They are organized by those that use matrix balancing, which is the most common, and
those that use list balancing. They are also further organized by whether they utilize both fitting
and drawing procedures, which seems to be the most common, or whether they only do fitting or
only do balancing. Note that most of these were developed at universities, with two exceptions
developed by consultants in Atlanta and San Diego. Other activity-based models now in practice
have used some version of these population synthesizers. Lets look at a few examples.
391
Page 39
Generalization
Balancing Procedures
Type
Controls
A priori weights
(seed distribution)
Contribution
coefficients
Multidimensional
Matrix (MORPC)
Row/column
totals
Initial matrix
Cell-row/column
incidence (0,1)
Table of categories
(ARC)
Column totals
Row/column
incidence (0,1)
List of individual
records (SANDAG)
Column totals
Initial individual
weight (row)
Row/column
coefficient (0)
39
Each subsequent method includes the previous one as a particular case and guarantees the
same result
Not every table of categories can be reduced to a matrix form!
Not every table of individual records can be reduced to table of categories!
392
Page 40
i 1
n 1
n2
n3
n4
n5
.
Control
HH size
2
3
i2
i 3
4+
i4
0-15
i 5
Person age
16-35 36-64
i6
i7
i 8
1
1
1
1
1
1
2
3
300
400
400
100
65+
200
250
2
2
2
650
HH
initial
weight
n
20
20
20
20
20
250
40
The list balancing procedure can be illustrated in the following way. The rows in the list are
individual HHs from the sample. Each HHs has an initial a priori weight that is normally equal to
all HHs but can be differentiated. The columns of the list table correspond to controls. HH-level
controls correspond to the total number of HH of certain type while person-level controls
correspond to total number of persons of certain type that we need to synthesize in the given
TAZ. The matrix contains contribution factors for each household to each control. Contribution
factors are 0,1 for household-level controls but it can be any number 0,1,2,3,4 for person-level
controls. The balancing procedures is applied to calculate HH weights that would match all
controls.
393
Page 41
41
This problem has a closed form formulation as the maximum entropy problem. We have
developed very efficient methods to find the unique solution. The essence of this formulation
(and that is what a good synthetic population is) is to calculate weights that would meet all
controls but also preserve the initial weights as much as possible (i.e. would use all HHs in the
sample in the most uniform way). Conventional matrix balancing (many of you are familiar) or
table balancing are particular cases of this general method.
394
Page 42
42
There are many advantages of the List Balancing procedure vs. matrix balancing applied in many
earlier population synthesizers. No reason to fight zero cells in joint distributions, they cannot be
utilized anyway. Can incorporate any HH-level and person-level controls naturally. Prepares
background for discretizing; no need in drawing HHs from the sample.
395
Page 43
43
Several useful extensions of this method have been recently introduced (MAG, BMC). One of
them allows for differential relaxation of controls. It solves a very frequent problem of nonconvergence if the controls are not fully consistent between themselves (which is a usual case in
practice for at least some TAZs in the region). In this formulation the balancing procedure would
find the best compromise solution. It can also naturally incorporate differential importance of
controls. Controls with high weight will be fully satisfied. Less important controls will be
satisfied to the extent possible.
396
Page 44
Importance of Entropy-Maximizing
Balancing vs. Simple IPF
Household
type
Initial
distribution
Linear IPF
Entropy-max
balancing
0 workers
20%
4%
12%
1 worker
30%
36%
25%
2 workers
40%
48%
47%
3 workers
10%
12%
16%
Control
average
1.4
1.8
1.8
44
This is not just a theoretical achievement. Balancing methods have a very important practical
impact on the results especially if structural changes in the population are expected. Consider a
zone with the initial seed (base year) distribution of HHs by number of workers as shown in the
second column. It corresponds to the average number of workers per HH of 1.4. Now, lets say
we need to restructure the distribution to meet a constraint of 1.8 labor force participation. A
nave IPF would result in an unrealistic distorted distribution shown in the third column. The
entropy-maximizing method yields a much more reasonable structural shift shown in column 4.
397
Page 45
Implementation Example 1:
45
This first example is a somewhat simple specification of the PopGen population synthesizer,
which has seen wide use in a number of locations. PopGen was developed at Arizona State. In
this implementation at the Baltimore Metropolitan Council (BMC), only three household
attributes are controlledfive levels of household size, four income levels, and four worker
categories. The spatial unit of control is the TAZ.
PopGen also has the ability to control person-level attributes, although this was not done here.
This is primarily because the synthetic population is being used to produce a joint distribution of
household size, income and workers for use in the trip generation stage of a 4-step modeling.
This is a preliminary step in BMCs long-term plans to develop an activity-based modeling
system.
398
Page 46
Implementation Example 1:
Control
Synthetic
% Difference
Household Size
1 person
2 persons
3 persons
4 persons
5 or more persons
Total
527,210
561,788
333,499
261,710
207,021
1,891,228
527,266
562,293
333,607
261,534
206,528
1,891,228
0.01%
0.09%
0.03%
-0.07%
-0.24%
0.00%
Household Income
< $11,800
$11,800 and < $26,000
$26,000 and < $44,200
$44,200
Total
190,133
284,592
378,935
1,037,528
1,891,188
189,005
283,968
378,845
1,039,410
1,891,228
-0.59%
-0.22%
-0.02%
0.18%
0.00%
46
This table shows the fit of the synthetic households to the control data, which as expected is quite
good. One thing to notice here is that the model is specified to be sensitive to lower-income
ranges, hence the first three income groupings are all below the median household income for the
region, with one category representing more than half of the households in the region. You may
also notice that categories which have the smallest total numbers (lowest income group, five-plus
person households) are also the most difficult to fit.
399
Page 47
Implementation Example 2:
47
In contrast to the Baltimore example, this example from Atlanta is a little more complex. It uses
six household control variables, having added the presence of children, two age of head
categories, and a family/non-family household indicator. There are no person control variables.
400
Page 48
Max % diff
Min % diff
Implementation Example 2:
Standard Dev.
48
Here are selected validation results from the ARC model. Although the model variables shown
are controlled at the TAZ level, validation statistics are shown at the Census tract level for
household income and at the PUMA level for household size. The comparison here was with a
backcasting exercise to see how well the model projected to a different year than the baseline.
This is a very useful approach to model validation.
The plots show maximum differences between synthesized and actual populations by the vertical
dashes, examples of which are circled in red for the first three categories.
Dots symbolize the minimum difference, and these are circled in blue for the first three
categories.
The longer horizontal lines, pointed out here in purple for the first three categories represent one
standard deviation about the mean. As you can see, there are some quite large deviations.
401
402
Page 49
Implementation Example 2:
Practical implications:
If uncontrolled variables are not directly used in the model these biases
are tolerable
Uncontrolled variable with a strong impact on travel demand should be
better re-specified and additional controls introduced
Activity-Based Modeling: Activity-Based Model Basics
49
With this ARC example, there are a couple important aspects. First, there is the bias that occurs
with uncontrolled variables. In this model, race was uncontrolled and backcasts revealed gross
inaccuracies in projecting the distribution of household by race.
Second, this example illustrates the bias that can occur in forecasts when the controls in the
forecast year are much more aggregate than in the base year. What happens is that the seed
preserves the distribution from the base year and the distribution of the uncontrolled variable
changes in the future year within the more aggregate category.
Other ARC results point out the problems of bias with uncontrolled categories. From the
practical perspective it is important to keep in mind how the problematic uncontrolled variable is
going to be used in the travel model. In general, important variables that have a strong impact on
travel demand have to be controlled in the population synthesis.
403
404
Page 50
Implementation Example 3:
50
The third example we have is from the SHRP2 C10A project in Jacksonville, Florida. This is the
same example we showed before in which we used a 3-dimensional joint distribution of
household size, income and workers as a control variable. This created 44 categories for this
single control variable. Also, in this model we have added explicit person control attributes for
age and gender. In addition, this project also modeled seasonal households, rather simply, by
their size and age, with control data obtained from the NHTS add-on survey for Florida. Adding
seasonal households is not something done in most regions, however, Florida cities these
households represent an important demand segment and appear in trip-based models. Seasonal
households typically do not work in the region, but exhibit travel behavior similar to other nonworking households.
405
Page 51
Implementation Example 3:
Control
Synthetic
% Difference
Household Size
1 person
2 persons
3 persons
4 or more persons
118,841
161,113
84,235
119,901
161,595
83,394
0.89%
0.30%
-1.00%
115,067
114,408
-0.57%
Total
479,255
479,298
0.01%
0-15 years
286,068
283,248
-0.99%
16-20 years
78,668
77,511
-1.47%
21-44 years
443,351
435,734
-1.72%
45-64 years
270,899
266,070
-1.78%
65+ years
123,868
122,237
-1.32%
1,202,855
1,184,800
-1.50%
Person Age
Total
51
Here are the goodness of fit statistics for this model. This shows that even with moderately
complex household and person specifications, good fit can be achieved.
406
Page 52
Implementation Example 4:
52
This example from San Diego illustrates an even more complex model specification in which we
control for both household and person types and have added dwelling types. This is the most
complex control structure for person attributes of the examples we have seen. By controlling for
race, this model is likely to avoid some of the statistical bias issues that the ARC model revealed
when race was included as an uncontrolled attribute. At the same time, this poses challenges for
forecasting race for future year populations.
Introduction of person-level controls is essential since certain demographic tendencies like
population aging are better described in terms of person distributions than in terms of household
distributions. This also creates a better linkage between the travel model and land-use and
demographic models.
Also note that there is a separate group quarters status, consisting of four categories. Modeling
group quarters populations is typically done separately from the non-group quarters population.
407
Group quarters residents are treated as if they belong to a one-person household, so their person
attributes, namely age and gender are typically the only attributes.
408
Page 53
Implementation Example 4:
Variable
Pop. Syn.
Census
1 person
24.2%
24.2%
-0.4%
1.5%
2 persons
32.3%
32.0%
0.8%
1.0%
3 persons
15.9%
16.1%
-1.8%
2.0%
4 or more
persons
27.7%
27.7%
-0.7%
3.3%
985,938
992,681
-0.6%
0.9%
Total HHs
Activity-Based Modeling: Activity-Based Model Basics
53
This slide shows another table of goodness of fit statistics for the household size distribution,
which look good. To be more precise, however, it is useful to see how closely the synthetic
population matches the Census (or other control source) at a finer geographic level. This map
shows Census Tracts in the region and how the distribution of synthetic persons ages 35 to 49
compare to the Census. A geographic analysis of this kind is recommended for validating a
population synthesizer.
409
Page 54
Types of Errors:
Under-representation of market sub-segments model may
be insensitive
Over-representation of market sub-segments model may
be too sensitive, or sensitive in unexpected ways
Mis-alignment of population with geography inaccurate
forecasts (trip lengths, mode shares, etc.)
Activity-Based Modeling: Activity-Based Model Basics
54
The outputs of a synthetic population are the inputs to all other model components. If these
inputs are inaccurate, forecasts and other analyses will also be inaccurate. We can generally
represent these as one of three types of errors:
If we are not specific enough in specifying the distribution of households and persons in the
model, then we are likely to under-represent certain market sub-segments, and the model may be
insensitive to key policy levers.
If we try to slice the distribution to finely, we may distort the distribution. This can lead to overrepresentation of market sub-segments, and the model may be too sensitive, or sensitive in
unexpected ways.
If we are inaccurate in where we place households, that can also lead to inaccurate forecasts in
travel behavior, just as they would with a trip-based model.
410
411
Page 55
55
Although the methods that have been developed are relatively robust, one cannot specify a
population generator naively. There are a few challenges to goodness of fit that one might need
to overcome.
First, there is the proper specification of the control attributes. Controlling for certain attributes
may distort the distributions of others. Several iterations of testing may be required before
determining the final set of control attributes.
Second, it is important to account for sparse data across both attribute and spatial dimensions.
Too many zero-valued cells in the seed matrix created from sample data could render the
balancing step infeasible. Solutions include: redefine or combine attribute categories (collapse
attribute levels to fewer groupings), and to expand the geographic unit to a more aggregate level.
412
Page 56
56
In practice, it is usually possible to maintain tight control on either household control targets or
person control targets, but not both simultaneously. Certain controls can be relaxed depending on
the importance (for modeling) and reliability of the data sources.
However, it is also possible to indirectly control person attributes at household level In practice,
the only person characteristics that are usually controlled for are age and gender. Gender should
work out without explicit controls. Age can be handled at the household level by using age of
head and presence and absence of people in certain age groups (i.e., very young children, school
age children, etc.)
413
Page 57
Sample population:
Re-use assume similar joint distributions of household and person
attributes in the horizon year appropriate in built-out areas
Consider changing demographics enrich the sample by adding HHs
from other geographic units that look more like your region in the future
(primarily, ethnicity or income mix)
57
One of the big questions that comes up when discussion population synthesis is where do we
get marginal control totals and sample populations for future years?
Control variables can be forecast using trend extrapolation methods, such as growth factoring, or
the outputs of a land use model, if available, may be used. Many regions will also have macrolevel regional socio-economic forecasts for future years, which can be use to grow the population
at the regional level.
The sample population, from which we draw household characteristics may be derived by either
assuming that the same types of households and correlation between control variables that exists
in the base year will also be present in the future year. For some regions, or sub-regions, this may
hold. For other regions, however, especially fast-growing metro area, this will most certainly not
be true.
414
One way to reflect changing demographics in the household sample is to borrow samples from
another areafor example another PUMA within the same regionthat looks more like the
way this subject area is expected to look in the future.
A third, emerging research area is household evolution modeling, which has the potential to
replace the more mechanical methods of population synthesis by modeling processes that reflect
actual human life courses. We will spend the remainder of this seminar exploring this topic.
415
Page 58
Household Evolution
Synthetic population for a base year is evolved to a
future year by simulating certain demographic
evolutionary processes
Base Year
aging
births
deaths
migration
student evolution
worker evolution
marriages
divorces
household formation
household dissolution
Activity-Based Modeling: Activity-Based Model Basics
58
Household evolution models have been under development at least since the early 1990s, but
have yet to be developed to the point where they can be used in practice. That said, they have a
lot theoretical appeal and are worth considering further.
In household evolution, a synthetic population for a base year is evolved to a future year by
simulating certain demographic evolutionary processes. These processes include: aging, births,
deaths, marriages, divorces, household formation, household dissolution and migration. There
are different methods applied for each one.
Of these, aging, births, deaths and migration are typically handled through well-accepted
demographic practices and are perhaps of the most immediate relevance.
The second group of itemsstudent and worker evolutionmay also be forecast, and there are
numerous good precedents available in the economics literature.
416
The third group of four itemsmarriages, divorces, household formation and dissolutionare a
bit more controversial and difficult to forecast. They may be important, however, for determining
the compositions of future households and their size.
417
Page 59
59
The effects of demographic shifts, such as the aging of the population, are not always
reflected in non-evolution forecasting processes.
Generate new HHs by a logical combination of features instead of re-using the same
small sample over and over
Most travel models and fail to account for in- and out-migration.
Socio-economic forecasts can be developed more methodically than applying naive
growth rates to marginal distributions.
Potential to provide more variables for dynamic travel models variables of change and
lagged variables.
Household evolution is consistent when integrated with land use models, which are also
usually evolution models.
418
419
Page 60
Source:
http://www-personal.umich.edu/~sdcamp/up504/cohort%20survivalexamplew08.pdf
60
The cohort survival method is a well-accepted method for aging a population based on fertility
and mortality rates for persons within certain age cohorts. In this example, 10-year intervals are
used, which is standard practice. Survival and birth rates are published by various state, federal
and international organizations. The U.S. Census also maintains birth rates, death rates in
migration rates for each state.
420
Page 61
Population Pyramids
61
This slide shows graphically how differences in fertility, mortality, and net migration affect the
distribution of the population by age groups. These types of graphs are called population
pyramids. As you can see, there are huge differences between the three nations depicted here (in
1995).
In fast-growing Kenya a vast majority of people are children. If we were to age this population,
those children will become adults and move up the population pyramid. If mortality and fertility
rates maintain the same levels as the 1990s, it is likely that the shape of the pyramid, which is
based on percentages, will not change.
If however fertility rates were to decline and/or mortality rates were to decline, then we might
see the pyramid taking on a shape more like that of the U.S. (center). Here we can see the baby
boom generations born between 1946 and 1964 about midway up the center of the graph. Their
children, the so-called boomlet are represented at the bottom of the graph. A third example
(right) is that of a zero-growth/declining population in Italy.
421
This is an example of a rapidly aging population where there are both much lower fertility and
mortality rates, compared with past generations. So, in Italy, the bulk of the population is moving
up the pyramid. This type of phenomenon may actually be happening in certain metro regions
within the U.S. particularly where there is a net zero or negative migration pattern.
422
Page 62
62
Cohort survival methods seem to be most applicable for producing control variables for persons.
It is not clear, however, how this should be reflected in the evolution of households. Obviously,
wed need additional information on how households are created, evolve, and dissolve, and how
this relates to the characteristics of individuals. To create synthetic households it seems
necessary to age households and persons at the micro level, but this requires predicting
individual births, deaths, migration and a host of other evolutionary processes. Lets consider
how this might play out at the micro-simulation.
423
Page 63
Person Evolution
63
Person evolution would seem fairly straight forward. Aging adds year to the population and, for
a given person, it seems straightforward to predict survival to the next cohort year. The challenge
is that mortality rates themselves have been changing and will continue to change with advances
in medicine.
Birth rates are also fairly straight forward in application; however, as we have witnessed, societal
norms have changed over time such that adults increasingly wait until later in life to have their
first child. In addition, predicting whether a given household will have children is not so simple
due to factors such as out-of-wedlock births, adoptions and foster care.
424
Page 64
Model
Region
64
In- and out-migration to a region is somewhat straight forward to predict at an aggregate level.
At the individual household level, however, there are several challenges. First, there is the
macro-economy which may spur sudden in or out-migration and can be quite unpredictable.
Second, there are challenges in identifying which households are likely to migrate outside of the
region, without some type of longitudinal tracking, because the evidence of their move is more
likely to show up in their new location than the one that they just left. Third, knowing the
attributes of households and person coming into a region may be a little easier to determine,
particularly from state sources; however, it is not clear how those trends will hold up in the
future.
425
Page 65
Student Evolution
Education
Determines education level of a person in the population
Challenge representing people who discontinue education
temporarily and later return to school
65
In order to get a more complete picture of how our synthetic population will evolve, some
models even go so far as to account for evolved student status and locations. There are of course
challenges, such as predicting persons who take courses part-time and go in and out of student
life. In addition, there are challenges associated with predicting whether certain household
members will start college/university, and whether that means they will stay local and remain in
the same household, or move out of the household into a group quarters situation or into a new
household (rent-apartment).
426
Page 66
Worker Evolution
Labor Participation
Predicts whether a person joins the labor force
Occupation Choice
Determines the occupational category of a person who is
predicted to join the labor force
Employee Income
This model predicts the income/eraning level of an employee
which can be used to estimate the household income
66
If we have student evolution, then of course we have to have worker evolution, which is even
more complicated. There are several types of models associated with worker evolution. First,
there are labor participation models, which simulate whether a person joins the labor force and at
what participation rate (full-time, part-time).
Second, there is occupation choice modeling, which seeks to determine the occupational
category of a person who is predicted to join the labor force.
Third, there are models that try to predict an individuals income, which may be aggregated up to
predict household income.
All three of these model types have long been the subject of economic analysis, so there are
plenty of examples to go by, some of which are probably overly complicated. The challenge is in
finding the right model for simulating individual decisions, and doing it in a household context,
where there are inter-individual decisions.
427
Page 67
Divorce/Household dissolution
Determines divorce decision of adults in the population
67
The last set of models we will mention are models of household formation and dissolution. These
are probably the least well understood of the models weve discussed here. How people come
together to form new households would seem to come down to a matching problem and is
complicated by the different circumstances that lead to persons living together, such as marriage,
cohabitation, and roommates (non-family households).
Perhaps even more murky is modeling household dissolution. While it is somewhat easier to
predict children growing up and leaving their parents home, based on a natural life-course
perspective, it is much more challenging and perhaps uncomfortable to try to predict divorces
and separation between adults who have been living together.
428
Page 68
68
There are several challenges that, to date, have prevented household evolution models from
making their way into common usage. These include choice of model form discrete choice,
regression, hazard-based; model sequencing and frequency of simulation; and the validation of
the framework and models pose a challenge in and of themselves. Thus, this is a very promising
research direction but some time is needed to bring it to the necessary level of maturity in
practice.
429
Page 69
69
430
Page 70
431
Page 71
February 2
February 23
March 15
April 5
April 26
May 17
June 7
June 28
July 19
August 9
August 30
September 20
71
Once again, here is the schedule for the webinar series. Our next webinar, three weeks from
today, will cover accessibility and the treatment of space
Thank you!
432
Page 72
72
433
John: A population synthesizer is a procedure that provides a linkage between a land-use model
and an activity-based model. It is not designed to shift households across zones, but rather create
a disaggregate population that respects input marginal distributions. However, there are hybrid
procedures possible, which can both shift distributions of households across zones and generate a
synthetic population. This is not typically done; instead, the household distributions are provided
by the land-use model and the population synthesizer respects those distributions.
435
436
Page 1
Activity-Based Modeling
Session 6: Accessibilities & Treatment of Space
437
Page 2
Acknowledgments
This presentation was prepared through the collaborative efforts of Resource
Systems Group, Inc. and Parsons Brinckerhoff.
Presenters
Joel Freedman, Kostas Goulias
Moderator
Stephen Lawe
Media Production
Bhargava Sana
Resource Systems Group and Parsons Brinckerhoff have developed these webinars
collaboratively, and we will be presenting each webinar together. Here is a list of the persons
involved in producing todays session.
Joel Freedman and Kostas Goulias are co-presenters. They were also primarily
responsible for preparing the material presented in this session.
Stephen Lawe is the session moderator.
Content development was also provided by John Gliebe and Rosella Picado. John
Bowman and Mark Bradley provided a review of the material.
Bhargava Sana was responsible for media production, including setting up and managing
the webinar presentation.
438
Page 3
February 2
February 23
March 15
April 5
April 26
May 16
June 7
June 28
July 19
August 9
August 30
September 20
3
For your reference, here is a list of all of the webinars topics and dates that have been planned.
As you can see, we will be presenting a different webinar every three weeks. Three weeks ago,
we covered the fifth topic in the seriesPopulation Synthesis and Household Evolution. This
session covered the creation of synthetic populations for use in an activity-based model
simulation.
Todays session is the third of nine technical webinars, where we will cover the details of
activity-based model design and implementation. In todays session, we will describe how
different activity-based model systems treat space, including both zone-based and parcel-based
systems. We will also cover how accessibilities are calculated and used in activity-based models,
and why they are important.
439
Page 4
Learning Outcomes
By the end of this session, you will be able to:
Describe why accessibilities are important in activitybased models
List important dimensions of accessibilities
Identify three main types of accessibilities
In todays session, we will be covering the treatment of space and accessibilities in activitybased models. At the end of this session you should be able to:
440
Page 5
Webinar Outline
Basic terminology
Need for accessibility measures
Important dimensions of accessibilities
Data needs
Types of accessibilities
Impact of accessibility on activity and travel choices
On-going research
Questions and answers
In this webinar, we will cover the different types of spatial treatments and accessibility measures
used in many activity-based models. We will discuss why accessibility measures are necessary,
what are the important components or dimensions of accessibility measures, and the impact of
accessibilities on activity and travel choices. We will talk about the treatment of space in
activity-based models, and implications for different treatments of space in terms of data needs
and computational burden. Finally, we will cover on-going research into accessibility measures
and provide time for questions and answers.
441
Page 6
Terminology
Skims
Density\Area type classification variables
Continuous accessibility variables
Mode choice logsum
Destination choice logsum
This slide shows basic terminology for that will be used in this session and other sessions.
Skims: Matrices containing level-of-service information (travel time, distance, cost) for each
origin-destination zone pair (or transit stop pair), for a given mode (drive-alone, shared-2, walk,
walk-transit, etc)) and time period (e.g. AM, midday, PM, etc).
Density\Area Type Classification: Each TAZ is classified as belonging to one of a set of
mutually exclusive, categorical area types, such as Central Business District (CBD), urban,
suburban, and rural. The classification typically considers zones size, employment, and
population, but does not consider level-of-service by mode. This is the most basic method of
accessibility classification and is not recommended for an activity-based model.
Continuous accessibility variables: The ease of travel to a destination or set of destinations, for
one or more household\person types, by one or more modes of travel, for one or more times of
day, for one or more activity purposes. Note key dimensions destinations, household\person
types, modes, time periods, activity purposes. Preferred for activity-based models because they
442
incorporate transportation system level of service and can be formulated to vary across these key
dimensions.
Mode Choice Log-sum: The natural log of the denominator of the mode choice model. A
measure which reflects accessibility by all modes of transport, according to the measured
attributes of those modes (e.g. time, cost, etc.), the traveler perceptions (or weights) for those
attributes, and the non-included attributes of the mode (reliability, safety, comfort, and traveler
biases). The measure essentially weights the contribution of each mode according to its
probability. A mode choice log-sum is useful when origin, destination, purpose, and time of
travel are known.
Destination Choice Log-sum: The natural log of the denominator of the destination choice
model. A measure which reflects accessibility by all modes of transport (see mode choice logsum, above) to all possible activity destinations. The measure essentially weights the
contribution of each destination according to its size, or attractiveness, and the mode choice
accessibility from the origin to the destination. A destination choice log-sum is useful when
origin, purpose, and time of travel are known. Note that destination is not known.
443
Page 7
Key Concepts
There is a growing recognition that accessibility affects a
wide range of travel dimensions:
Route choice
Mode choice (including car occupancy)
Time-of-day choice
Tour and stop destination choice
Daily activity pattern generation
Car ownership choice
Workplace and Residential location choice
There is a growing recognition that transport accessibility to land-uses affects a wide range of
travel dimensions. One of the primary motivations for adopting an activity based model is so that
accessibility provided by the transportation system can be treated in a consistent and
comprehensive manner on travel decisions and thereby improve policy analysis. Here is a listing
of how accessibilities might affect travel behavior:
444
accessible destinations typically have a higher probability of selection, all else being
equal
Daily activity pattern generation: Generally travelers residing in more accessible
locations tend to make more tours on average, and less stops per tour, than travelers who
reside in less accessible destinations
Car ownership choice: Households located in areas with better transit and non-motorized
accessibilities tend to own fewer cars, on average, than households in less accessible
locations
Workplace and Residential location choice: areas with higher accessibility tend to be
more attractive places to work and to move to. A classic concept in economics is the bidrent curve, which shows that higher-accessible locations tend to have higher land-values
due to the greater demand for those locations due to their accessibility
Accessibilities are used to represent the influences of transport policy, land-use policy,
development patterns, geographical constraints, and congestion on these choices.
445
Page 8
Accessibilities
Synthetic
Population
Downward
Integrity:
Choices made
in higher
models affect
choices made
in lower
models
Mobility
Choices
Model
Inputs
Long-Term
Choices
Daily
Activity
Patterns
Trip
Assignment
Upward
Integrity:
Expected
utility of
making
choices in
lower models
affect choices
made in higher
models
Model
Outputs
Activity-Based Modeling: Accessibilities & Treatment of
Space
This diagram was presented in Webinar 4, but it is important to review this concept before we
continue. While activity-based models can vary in structure, this diagram shows the location of
tour and trip detail choices (tour mode, primary destination, intermediate stop location and trip
mode) in a typical model stream. As the model system progresses, travelers make decisions:
whether to travel, where to go, how many stops to make, what mode to choose, and so on. Earlier
decisions influence and constrain the decisions made later; for example, the number of vehicles
owned, modeled in the auto ownership (mobility) model, influences the number of tours and the
mode used on each tour. The mode used for the tour then influences the location of stops on the
tour, and so on. This is referred to as downward vertical integrity.
Activity-based models also use information from models that are lower in the model chain to
inform the choices made by decision-makers in upper-level models. This information typically
takes the form of accessibilities that are based upon all of the information that is relevant for a
lower level choice. For example, a mode choice log-sum, which reflects accessibility by all
modes of transport, can be used to inform the choice of destination for the tour or stop. This is
446
referred to as upward vertical integrity. The upward integrity of the model system is
represented via accessibility terms.
447
Page 9
Defining Accessibility
Origin-Destination Based:
Origin-Based
How many shopping opportunities can I get to from my home?
How many shopping opportunities can I get to from my home by transit at 9:30
A.M.?
How accessible is my home to shopping opportunities at 9:30 A.M. by all modes
of transport?
How accessible is my home to shopping opportunities throughout the day by all
modes of transport?
There are a number of ways to define accessibility, and we will be exploring them all in this
webinar. There are two common types of accessibility measures. One is based upon a given
origin-destination pair. It quantifies the time and\or cost of travel between two points in the
region (for example, home and work). The slide starts with asking a nave question: how long
will it take to get from home to work? However, this is a nave question because it is missing two
key aspects of travel by what mode, and at what time of day?
The next question is much more specific how long will it take to get from home to work by
transit at 9:30 AM? Now that the question is asked more specifically, it can be answered with
some fairly straight-forward calculations, as we shall see later in this webinar. The question can
also be asked in a more general sense what is the accessibility of home to work throughout the
day by all modes of travel? Calculating accessibility across multiple time periods and/or modes
is possible, and we will cover this concept later in the webinar.
Another type of accessibility measure is an origin-based accessibility. A nave question might be
How many shopping opportunities can I get to from my home? But again, without specifying
448
a mode and/or time, this is difficult to answer. However, it demonstrates that origin-based
accessibilities are typically specified for a certain activity purpose; in this case, shopping, from a
certain origin, in this case, the home, to all destinations. The next question includes both a mode
and a time. The last two questions summarize accessibility across all modes and or times. Again,
there are different ways of calculating these accessibilities, as we shall see.
449
Page 10
Population
Long-Term
Mobility
Modes
Space
Activity-Based Modeling: Accessibilities & Treatment of
Space
Activities
Tour Patterns
Time
10
Hopefully you saw this diagram in the last webinar on population synthesis. It serves as a
backdrop for describing the relationships between key design elements in activity-based
modeling. These elements include: defining the population, modeling long-term and mobilityrelated choices, defining activity types, defining modes, defining tour patterns and an entire daypattern elements, as well as the treatment of space and accessibility and treatment of time. The
elements that relate most directly to accessibilities are shown in red. They include the definition
of space, of time, the modes that the model considers, and the types of activities that people
undertake. Accessibility measures are ways to quantify various amounts of these elements.
Accessibilities influence all facets of an activity-based model system. After all, one of the key
purposes of a travel demand model is to measure the affects of transport and land-use changes on
travel demand and system performance. Accessibility variables measure transport and land-use
system performance.
450
451
Page 11
Transit analysis
Capture the affects of transit service changes on auto ownership
11
Here are some example policy effects that are important to capture in an activity-based model,
and in fact are primary motivations for agencies to adopt activity-based models.
Peak spreading occurs as a result of increasing congestion in a region. Time-of-day choice
models capture the effects of congestion on the choice of when to travel. These affects are
measured in terms of the accessibility provided between the origin and destination of the tour at
different times of day. All else being equal, more congested times of day are less desirable times
to travel. We will see exactly how this works in the session on tour scheduling on July 19.
Transit analysis can be improved by consideration of transit accessibilities on auto ownership.
This is common in some of the more advanced trip-based models as well. However, activitybased models can consider the accessibility for specific workers based upon where they live and
work in the household auto ownership decision. We will see exactly how such variables are
included in auto ownership models in the Session #7 on long-term and mobility models on June
7.
452
453
Page 12
Frequency
25%
20%
Work
15%
Maintenance
10%
Discretionary
At-Work
5%
0%
1
11
13
15
17
19
21
23
25
27
29
Distance in Miles
Source: 2005 San Diego Household Survey Data, San Diego Association of Governments
12
This slide shows tour length frequency distributions by tour purpose (distance is one-way
between tour origin and primary destination). It demonstrates the importance of considering tour
purpose in accessibilities and how different tour purposes have different sensitivities to distance.
For example, work tours tend to be much longer than shopping (maintenance) tours.
454
Page 13
Percent
25%
Toll: All
20%
Non-Toll: All
15%
Toll: Work
10%
Non-Toll: Work
5%
1:00 AM
2:00 AM
3:00 AM
4:00 AM
5:00 AM
6:00 AM
7:00 AM
8:00 AM
9:00 AM
10:00 AM
11:00 AM
12:00 PM
1:00 PM
2:00 PM
3:00 PM
4:00 PM
5:00 PM
6:00 PM
7:00 PM
8:00 PM
9:00 PM
10:00 PM
11:00 PM
12:00 AM
0%
Source: 2011 Atlanta Regional Commission Household Survey Data, Atlanta Regional Commission
13
This slide shows trip departure hour for toll and non-toll trips. Note that toll trips are more
peaked; the increased accessibility (faster travel times) provided by toll lanes provides greater
opportunity for travelers to depart in the peak period; much more so for work trips than other
trips.
455
Page 14
Tours Per
Tours Per
Stops Per
Household Person
Tour
4.34
2.49
4.30
2.40
4.77
2.40
5.33
2.36
5.00
2.17
5.88
2.32
5.49
2.24
5.35
2.28
0.57
0.55
0.60
0.62
0.61
0.58
0.70
0.62
Source: 2000 Atlanta Regional Commission Household Survey Data, Atlanta Regional Commission
Tours per household are lower in urban areas because household sizes tend to
be smaller
Tours per person tends to be higher in urban areas due (in part) to increased
household accessibilities
Stops per tour tend to be much higher for rural areas due to decreased
accessibility around the home
Activity-Based Modeling: Accessibilities & Treatment of
Space
14
This slide shows a simple tabulation of tour and stop rates by area type from an older Atlanta
household survey. Tours per household tends to decrease with respect to density because
household size tends to decrease with respect to density. However, tours per person tends to
increase with respect to density, due in part to increased accessibility in more dense parts of the
region. Conversely, stops per tour tends to be relatively flat, but are typically greater in more
rural areas, where persons who live in those areas typically link together multiple trips due to the
lower accessibility to activities around the home.
Notes: Most recently-developed activity-based models utilize continuous measurements of
accessibility as opposed to area-type classifications (or dummies) as shown in this slide.
However, such definitions are helpful to visualize differences in the data. This webinar will
address the calculation of continuous accessibility measures in later slides.
456
Page 15
Modes
Time periods
To reflect different levels of congestion and supply
Peak\off-peak or more fine-grained
Activity purposes
Different land-use types are important for different activity purposes
Typical activity purposes have been listed in other presentations (mandatory,
maintenance, discretionary)
15
As can be seen from the previous slides, early measures of accessibility were rather limited.
They considered only land-use data, but not transport accessibility (as in the case of ARCs area
type measure). Or they considered transport accessibility, but only in a limited fashion (as in the
case of PAGs transit buffer). There are a number of components or dimensions of accessibility
measures of which activity-based models are currently taking advantage. They include:
458
Page 16
Land-use data
Synthetic population
Employment by category
Parking supply, cost
4D variables intersection density, sidewalks, topology, etc.
Survey data
Household survey data, properly weighted and expanded
Required to estimate size term, parameters
16
This slide shows the typical data required for calculation of accessibility measures. They include:
The zoning system used for representation of space, which also defines access to the
network and is typically the basic unit of analysis for level-of-service representation.
Transport networks including highway and transit representation. Networks are skimmed
to create level-of-service matrices, typically by mode and time-of-day. For example,
highway skims might be created for single-occupant vehicles and multi-occupant
vehicles, and further segmented by free and pay alternatives. Transit skims can be created
for local bus versus premium transit services. The skims are typically created for different
time periods, such as AM peak, midday, PM peak and evening. Finally, networks can be
skimmed by market segment, such as income, in order to reflect willingness-to-pay
differences between income groups (for example, low income travelers may be unwilling
to pay for higher price toll and transit alternatives).
Land-use data is required in order to represent the opportunities available to travelers,
such as jobs or retail employment for shopping. The more land-use categories that are
459
available, the better models will capture differences in accessibilities for various types of
out-of-home opportunities, but this comes at the cost of maintaining and forecasting
many employment categories. This is why land-use models are often desired. Parking
supply and cost is also useful to represent the increased cost of auto for certain areas such
as CBDs, and urban form or 4D variables are also helpful to describe the ease of nonmotorized travel.
Finally, survey data is necessary to measure the influence of accessibilities on travel
behavior; model parameters are estimated using household and other survey data.
460
Page 17
Spatial Systems
Spatial Representation
Diagram
Zones
Already exists for most MPOs
The most aggregation error, particularly for non-motorized and transit modes
Sub-zones
Created by buffering around transit lines, stops
Improves representation of walk-transit, but can result in inconsistent transit
times between buffers and skims
Doesnt help with non-motorized representation (intra-zonals)
Micro-zones
Created by sub-dividing zones (7-10:1)
Best representation of transit accessibility when coupled with stop-stop
skims
Improved representation of non-motorized time
Parcels
Created via parcel database
Best representation of short distances and travel times
Precise measurement of size and neighborhood effects
Improves representation of walk-transit, but can result in inconsistencies
between walk times and skims
17
Now we will begin to explore the data required for accessibilities mentioned on the previous
overview slide. This slide shows commonly-used spatial systems in activity-based models, and
describes some trade-offs between them. They will be described in more detail in following
slides.
TAZ Advantages:
Readily available
Generally easy to estimate land-uses
TAZ Disadvantages:
Aggregation bias w.r.t intra-zonal and close-in travel, particularly for transit
access/egress and non-motorized travel
Micro-zones Advantages:
Micro-zones Disadvantages:
Parcels Advantages:
Parcel Disadvantages:
462
Page 18
Metropolitan
Transportation
Commission
Transportation
Analysis Zones
18
This map shows the zone system used in the Metropolitan Transportation Commission activitybased model, which covers the nine-county San Francisco Bay Area. There are approximately
2,000 TAZs, ranging from very small (<.25 acres) in CBD to very large (over 100 sq. miles) in
outlying counties.
Note that MTC is currently updating their TAZ system, and will likely be moving to either a
micro-zone or parcel representation, specifically because the large zones cause problems in the
representation of non-motorized accessibility both in terms of walking within a zone, as well as
the walk-accessibility to transit.
463
Page 19
Metropolitan
Transportation
Commission
Sub-zone Map
19
Several activity-based models (Columbus, MTC, ARC, Chicago) currently use transit sub-zones
in their representation of transit accessibility. This method is consistent with many trip-based
models, in which the transit network is overlayed on the TAZ system, and a walkability buffer
is created around transit stops or routes to identify the portion of the TAZ within short or long
walk to transit (typically 1/3 mile is used for short walk, 2/3 mile for long walk). Each TAZ is
then split proportionally into short, long, and no-walk to transit shares, effectively tripling the
number of potential activity locations (although certain zones are 100% short walk or 100% nowalk, which results in less than zones3 locations.
Zone-based skims are still used to calculate transit times and costs between zone-pairs. However,
the walk-access time is then segmented by the portion of the zone in which the origin/destination
pair resides. For example, lets assume that a particular a trip origin is in a large zone, but the
zone is split into short walk (25% of the area), long walk (50% of the area) and no-walk (25% of
the area). Then, we draw a random number and determine which segment the trip is in. If the trip
is in the short-walk portion of the area, the potentially long time that was calculated in the transit
464
skims (due to the long centroid connector) is over-ridden by a short walk time, equivalent to 1/6
of a mile (since the walk buffer for short walk is 1/3 mile, the average walk time would be of
that buffer).
465
Page 20
Micro-zones
(Maricopa Association of Governments)
20
466
Page 21
Parcels
(Sacramento Area
Council of
Governments)
21
A number of activity-based models rely upon parcels as one of the representations of space, as
shown. Parcel-based systems offer the most accurate measurement of non-motorized (walk and
bike) times for close-together pairs of parcels. Parcel systems also allow one to more accurately
measure urban design variables such as land-use mix.
Typically, parcel based systems are also used to represent transit accessibility. Each parcel is
coded to its closest transit stop for each transit sub-mode, and that time replaces the skimmed
walk-transit time. It is mainly suitable for relatively simple transit systems, because in more
complex transit systems with several possible competing routes between zones, the parcel-tostop time may not be consistent with the skimmed zone-zone in-vehicle and out-of-vehicle times
or mode combinations used in those skims. For example, the zone-zone transit skim may
represent an LRT path with a bus transfer at each end. Since skims do not represent the order of
the path, it may not be possible to identify which transit stops to use for the trip. It may be
possible to use the method described earlier for micro-zones, to combine the access time from
467
each parcel to different stops with stop-to-stop transit skims. However, as we shall see later, this
may impose a huge computational burden on the model system.
Parcels require a parcel inventory\database, a method to deal with how parcels may change over
time, and a method to allocate base and future-year employment data to parcels. Methods are
being developed to use templates for future year new development areas, so that the land use
and accessibility measures will be accurate for that type of neighborhood, without having to
specify exactly where each new parcel and local street will be located. All-streets networks may
also be used with both micro-zones and parcels to better represent the influence of physical
barriers on walk time.
468
Page 22
Diagram
Area Type
The most aggregation error
Does not consider network level-of-service, by time period
Discrete, lumpy
Buffer Variables
Example: Number of jobs within 30 minutes transit service
Can be extended by time period
Typically limited to one mode, employment type
22
There are a number of different ways to calculate accessibility measures, as we shall see in this
webinar. They each have advantages and disadvantages, and specific uses within the activitybased model system.
The simplest form of accessibility measure is the flat area-type measure, which classifies each
zone as CBD, urban commercial, urban residential, suburban commercial, suburban residential,
and rural (or some variation thereof). The advantage of this measure is that it is easy to calculate.
But there are a number of disadvantages, the biggest of which is that it completely ignores
network level-of-service. Therefore any changes to the network (such as introduction of transit
service or toll lanes) would have no effect on this measure. For this reason, its use has been
limited to only the earliest activity-based models.
Buffer variables typically search around a zone and measure the accessibility of the zone to
employment within a certain range of travel time. For example, a buffer variable might measure
the number of retail jobs within 30 minutes of transit service. These measures can be calculated
for specific modes and time periods, and do take into account network level-of-service. But they
469
are specific to a certain mode or time period, which can also be a disadvantage. For example,
maybe I want to measure the time by both walking and biking. How do I correctly weight the
influence of each in travel time?
That is where mode choice log-sums provide a key advantage. They measure accessibility across
multiple modes, applying appropriate weights on the components of time and cost associated
with each mode, according to traveler preferences (as revealed by the mode choice model).
However, when time of travel is not known, a simple mode choice log-sum is insensitive to
changes in transport conditions that vary by time of day, so either a joint mode-time log-sum or a
mode log-sum that uses a probabilistically simulated time of day is needed. Mode choice logsums are specific to an origin-destination pair and are usually case-specific; that is, they are
calculated for a particular choice situation taking into consideration the characteristics of the
household, person and tour.
An extension of the mode choice log-sum is the destination choice log-sum, which measures
accessibilities across multiple modes, from a specific origin to all destinations. Destination
choice log-sums take into account the amount of activity at each destination. That activity is
often represented as purpose-specific. As with the mode choice log-sum, when time of travel is
not known, the log-sum should be enhanced to account for variations in conditions that occur
across the possible times of day.
Now well take a closer look at how each accessibility measure is calculated.
470
Page 23
23
Here we show how area types were calculated for the Atlanta region for their activity-based
model, originally developed in the early 2000s. Population and employment densities are
calculated and used to code each TAZ to one of seven area type categories. Similar methods
have been used in other regions, and are common in trip-based travel models. Area type
classifications offer advantages and disadvantages. In our opinion the disadvantages outweigh
the advantages and therefore are typically not used in activity-based models.
The advantages of area type classifications are as follows:
472
Page 24
Atlanta Regional
Commission
Area Type Map
Easy to calculate and use
However, does not consider:
transport accessibility
purpose of travel
temporal affects
And is somewhat lumpy
Activity-Based Modeling: Accessibilities & Treatment of
Space
24
Here is a map of the Atlanta Regional Commission area type classification system, focusing on
the more populated part of the 20-county ARC region. As you can see, the area types are
somewhat lumpy, with clear pockets of dense and less dense area types, which is somewhat
exacerbated by the use of large transportation analysis zones as the spatial unit in the model
(more on this later). However, the ARC activity-based model system was one of the first
activity-based models to be developed in the United States, and is therefore more aggregate than
many of the more recent models developed, as we shall see.
473
Page 25
Buffer Variables
n
25
Accessibility measures such as the one shown on this slide use buffers to count or aggregate the
number of opportunities within a certain distance or threshold around the TAZ given some modal
level-of-service. The example shows a count of the number of retail jobs within 30 minutes
transit time to measure transit accessibility. This is an improvement over the dummy area type
variable shown in the previous slide, because it takes into account network level of service.
However, it still has a number of disadvantages, such as cliff effects (that zones are either within
or outside the buffer), and that only one mode, time period, and land-use category are
considered.
474
Page 26
Pima
Association of
Governments
Transit Buffer
Map
26
Here is a transit buffer map created for the Pima Association of Governments. This was actually
used in the auto ownership model component of the four-step travel model. It uses a threshold of
20 minutes travel time by transit and is strongly correlated with transit service in the urban core
and radiating out from the core along major transit corridors. The lumpy effect can be easily seen
here, particularly due to the use of relatively large zones. If this variable were to be used in auto
ownership, a household on an edge of a zone may have a very different probability of owning 0vehicles than a household just on the other edge of the zone, as shown.
475
Page 27
Networks in accessibilities:
Why are networks important?
Reflecting transport options in accessibilities
How will construction of new highway\transit infrastructure affect
the propensity to travel?
How would toll roads\congestion pricing affect travel behavior?
27
Network representation is very important in accessibility calculations, since the main goal of
travel demand models is to measure the impact of transportation infrastructure on travel
behavior. As shown previously, more simplistic accessibility measures only consider density, or
only consider one mode of transportation. However, our models should be multi-modal in nature.
If a simplistic accessibility measure is used, only changes in the mode that is considered will
influence travel behavior. For example, if the model does not explicitly consider toll choice, it is
likely that the scheduling of toll trips will not be influenced by the faster travel times in the peak
period provided by the toll facility. However, we cannot just average the toll and non-toll travel
times, because not all travelers are willing to pay for the toll facility. Similarly, we want to
include the influence of transit, but we cant just average the auto times and the transit times.
Some travelers may be more or less willing to take transit than other travelers, and there are
many attributes of the choice that arent reflected just in time and cost of each mode.
So the bottom line is that we want a measurement of accessibility that is multi-modal, but takes
into account the times and costs of the various alternatives, weighted by the traveler perceptions
476
of those times and costs, as well as traveler characteristics, and all of the other attributes of those
modes that arent explicitly included in time, cost, or other traveler characteristics. That way, we
can calculate a true average accessibility across all the modes. The good news is that there is a
very convenient measure that does this. It is the mode choice log-sum. Well now see how this
works.
477
Page 28
Modal Utilities
Utility is the weighted sum of the attributes of the mode
Weights can vary depending on attributes of traveler, tour\trip purpose
Mode-specific constant quantify non-included attributes of mode
Utilityauto =
+
+
+
-0.025
-0.002
-0.050
1.500
*
*
*
*
in-vehicle time
parking cost + operating cost (cents)
access time and egress time
autos > drivers
+
+
+
+
+
+
-0.025
-0.050
-0.050
-0.063
-0.002
0.795
-1.708
*
*
*
*
*
*
in-vehicle time
access time and egress time
first wait time
transfer wait time
transit fare
16 < age < 24
Utilitytransit =
28
In a mode choice model, a utility is calculated as the weighted sum of the attributes of each
alternative. Utility is a numeric value that is assigned to each alternative that represents how
useful the alternative is to the decision-maker. The weights, or parameters, used for each
component of time and cost are calculated based upon observed data, typically household survey
and transit on-board survey data. The utility equations shown on this slide are based upon models
estimated for the San Diego association of governments; however, the alternatives and
parameters shown are a subset of the actual models, for the sake of simplicity. A few things to
note about the utility equation:
Some of the variables are based upon the origin and destination of the trip (such as invehicle time, walk time, wait time, and fare)
Some of the variables are based upon either the origin or the destination of the trip (such
as parking cost)
Some of the variables are based upon the characteristics of the traveler, such as the
number of autos chosen for the household compared to the number of drivers in the
478
household in the auto utility, as well as the age of the traveler in the transit utility. Further
note that the number of autos for the household is the outcome of a previous model, while
the number of drivers in the household is a field in the synthetic population. One of the
critical advantages of an activity-based model is that the utility equations can take into
account the specific characteristics of the traveler including person and household
characteristics.
Alternative-specific constants reflect the non-included attributes of the mode. They
reflect the probability of choosing the mode, all else being equal. In other words, they
reflect the probability of the mode if the times and costs of the alternatives were exactly
the same.
So, here we can see that the mode choice model takes into account times and costs of the modes,
weighted by the perceptions of the travelers, as well as the characteristics of those travelers, and
the influence of land-use characteristics, such as parking cost, on the mode choice. This is
exactly what we want for a good accessibility measure.
479
Page 29
Nonmotorized
Auto
Drive alone
Shared ride
2
Shared ride
3+
Walk(9)
GP(1)
GP(3)
GP(6)
Bike(10)
Pay(2)
HOV(4)
Pay(5)
Transit
Walk
access
PNR
access
School
Bus(26)
KNR
access
Local
bus(11)
Local
bus(16)
Local
bus(21)
HOV(7)
Express
bus(12)
Express
bus(17)
Express
bus(22)
Pay(8)
BRT(13)
BRT(18)
BRT(23)
LRT(14)
LRT(19)
LRT(24)
Commuter
rail(15)
Commuter
rail(20)
Commuter
rail(25)
29
Here we see all of the modes represented in the San Diego mode choice model. Auto modes
include drive-alone, shared-2 and shared 3+, with sub-options of general purpose, highoccupancy vehicle, and pay (toll) alternatives. Walk and bike are both represented explicitly, as
well as five separate transit line-haul modes (local bus, express bus, bus rapid transit, light-rail
transit, and commuter rail) and three access modes (walk, park-and-ride, and kiss-and-ride).
School bus is also an option for school tours. This diagram will help you understand the plots on
the next few slides. We will discuss various mode choice formulations and
advantages\disadvantages of the various options in webinar 10 on August 9.
480
Page 30
Ui
Pi
Ui
e
ln[
]
iI
30
The way that we convert these utilities into a single accessibility is to take the log-sum of the
mode choice model. The denominator of the mode choice model is the sum of the exponentiated
utilities of the model. If we calculate the natural log of the denominator of the model, we are
essentially converting this sum back into a utility, but it is a utility across all modes of travel. In
other words, it is a multi-modal accessibility, taking into account the times and costs of all the
modes, weighted by traveler perceptions, and traveler characteristics, and non-included
attributes.
A log-sum can be representative of certain conditions. For example, a non-motorized log-sum
can be created, where only walk and bike modes are available, or a walk-transit log-sum can be
created where only transit modes are available. Log-sums can also be created for a specific
decision-maker; for example, a mode choice log-sum for persons in 0-auto households would
reflect transit modes to a greater extent than a mode choice log-sum for 1+ auto households.
Log-sums are often used as variables in models to ensure upward integrity of the model
system; for example, the impact of accessibilities by all modes of travel on destination choice is
481
reflected through the use of the mode choice model log-sum in the destination choice model.
This will be explained further in the webinar on destination choice.
482
Page 31
31
Mode choice log-sums are inherently spatial. That is, they are relevant for specific origindestination pairs. When mode choice log-sums are created for an entire region, they are typically
stored in matrices. In order to visualize them, one must select either a row of the matrix (a set of
log-sums from a given origin to all destinations) or a column of the matrix (a set of log-sums
from all origins to a given destination). Here we see a plot of mode choice log-sums for San
Diego, where a column of the log-sum matrix was selected the column selected corresponds to
a micro-zone in downtown San Diego, as denoted by the green asterisk. These log-sums focus
only on auto modes in other words, only the auto modes were active in the mode choice model
that created this map, as shown on the right. You can see that the darker shaded areas of the map
are closer to downtown e.g., more accessible. And the lighter areas are less accessible to
downtown. Also, there are dark bands around freeways, indicating the accessibility that these
faster facilities provide to downtown.
Areas in grey are micro-zones that do not have any population or households in them. For the
sake of computation efficiency, log-sums were not calculated from these micro-zones.
483
484
Page 32
32
This plot shows the mode choice log-sum taken across walk-transit modes, for all micro-zones to
downtown. Contrast this plot with the auto accessibility plot, and notice that the walk-transit logsums are quite different from the auto accessibilities. The walk-transit accessibilities closely
follow transit routes. There is also considerable variation in accessibilities as one moves away
from the main routes, reflecting greater walk access and egress times required to use the routes
(as per the utility equations shown on the previous slides). Households located close to transit
routes will have increased accessibility as a function of the accessibility provided by transit. This
is a useful variable in measuring the impact of transit accessibility on household auto ownership,
and the walk-transit mode choice log-sum is often used as an explanatory variable in auto
ownership for this reason.
485
Page 33
Nonmotorized
Walk(9)
Bike(10)
33
Here is a close-up of the non-motorized mode choice log-sum from all micro-zones to a
downtown San Diego micro-zone. Note that the accessibility calculations include both walk and
bicycle modes. The inclusion of bicycle mode tends to flatten the accessibility calculation up to
the maximum bicycle distance. In this model, the maximum bicycle distance was set to 8 miles,
so the micro-zones that are further than 8 miles from the downtown zone are shown in grey. The
maximum walk distance was set as 3 miles, so micro-zones that are within 3 miles of the
downtown zone are much darker, reflecting the influence of the availability of walking as a
modal option.
This points out another important aspect of accessibility calculations, and models in general the
use of thresholds can have significant impacts on the explanatory variables, and therefore have
significant impacts on the predictions of the model system. It isnt really true that a micro-zone
that is 3.0001 miles from downtown has no walk-accessibility while a micro-zone that is 2.9999
miles from downtown does. One of the important aspects of model development is exploring the
impact of such thresholds. In the case of San Diego, the utility of walking 3 miles is already quite
486
low an hour of walking time is required; therefore, it was decided that setting a threshold of 3
miles would not be detrimental in the model system.
487
Page 34
34
The maps above show how certain modes can be turned on or off to calculate mode choice
accessibilities by a sub-set of modes. Another technique is to calculate log-sums by market
segment. This slide shows log-sums for 0-auto households versus auto sufficient households.
Again, it is for the San Diego region, from all micro-zones to the downtown zone. The plot on
the top left is for 0-auto households. These log-sums are influenced more by transit and nonmotorized modes. Why? Because in the mode choice model, the alternative-specific constants for
0-auto households are higher for transit and non-motorized modes and more negative for auto
modes (since they dont have a personal vehicle available). The plot on the top right shows the
mode choice log-sums for auto sufficient households to downtown. Auto sufficient households
are those households who have at least one vehicle for every driving age household member. At
first glance, it doesnt appear that the log-sums are very different for 0-auto households and auto
sufficient households. However, if you subtract the 0-auto log-sums from the auto-sufficient logsums, and plot the results (advance to show animation), then you will see that the difference is
the influence of the transit accessibilities. In other words, the difference plot follows the walktransit accessibility plot closely.
488
489
Page 35
Dimensions of accessibilities:
Tour/trip purpose
Different tour & trip purposes are attracted to different
types of land-uses
35
So far weve covered the treatment of space in terms of the zoning system used, and different
ways to measure accessibility in terms of network characteristics such as auto, non-motorized,
and transit mode alternatives and their characteristics. Now lets look at how accessibilities can
take into account the purpose of travel.
Obviously different types of travel are attracted to different types of land-uses. For example,
workers are attracted to employment, either total employment or industry type depending on the
workers occupation. So an accessibility measure that is going to be used for a residential
location choice model might want to consider accessibility to the types of jobs that workers can
work in based upon their occupation. Shopping trips are attracted to retail space or employment.
Eating out trips are attracted to restaurant employment. And so on. The possibilities are only
limited by the availability of good base-year data on the location of different types of
employment, and the ability of the agency to forecast the location of the employment by type
into the future. But the bottom line is that we want to have the access to those types of activities
represented in our measure of accessibility.
490
491
Page 36
Accessibility to activities:
Destination choice model
Choice
Uj = LS * mode_choice_logsumij + * X
+ ln(retail_emp + service_emp * service_emp)
Quantity variables (size term)
Activity-Based Modeling: Accessibilities & Treatment of
Space
36
The model that is used to determine the location of out-of-home activities in an activity-based
model system is the destination choice model. This slide shows the model form. Each zone (or
micro-zone, or parcel) is a potential activity location. In most activity-based models, the measure
of accessibility or impedance between the origin zone and potential destination zones is the mode
choice log-sum, as described in previous slides. The mode choice log-sum used in the model is
based upon the traveler making the decision of where to go, as well as the purpose and time of
travel. The attractiveness of each potential destination is based upon its size, as shown in the
logged expression. In this case, the example utility is for shopping. So, the relevant variables of
zonal attractiveness are retail and service employment. The reason why these terms are logged is
so that all else being equal, the number of tours or trips attracted to the zone will be proportional
to its size. When a probability is computed, the utility is exponentiated, so logging the size term
makes the probability of selection proportional to the size of the zone.
One can think of the destination choice model as a nested, or simultaneous, destination and mode
choice model.
492
493
Page 37
Accessibility to activities:
Destination choice logsum
Choice
Z I U p ,i
ln e
z 1 i 1
37
Because the destination choice model takes into account the modal alternatives between the
origin and all potential destinations, as well as the land-use or activity opportunities available in
each destination, it is a great way to calculate accessibility. In this case, the log-sum of the
denominator of the destination choice model is taken. Another way to write the destination
choice accessibility is shown on this slide. This accessibility measure is origin-zone based. It is
the accessibility to all potential activities, weighted by the impedance to the activity. It is often
computed for each specific activity purpose separately; for example, shopping, eating out or
other discretionary just as there are different destination choice models for each activity
purpose.
494
Page 38
38
Destination choice log-sums are origin-based, because they are computed across all destinations.
In other words, origin-based log-sums are represented as vectors of data, which can be stored in a
table with one value for each origin zone. Each vector can be easily plotted, as shown here. This
plot shows auto access to non-mandatory destinations. This uses the auto mode as the measure of
impedance, and non-mandatory purpose for the size term in the destination choice model. Areas
that are close to the main employment hubs in San Diego are darker, because they are more
accessible. Think of this plot as showing how many non-mandatory destinations can I get to by
car from my zone compared to some other zone.
495
Page 39
39
This plot shows transit access to non-mandatory destinations. This uses the transit mode as the
measure of impedance, and non-mandatory purpose for the size term in the destination choice
model. Areas that are close to transit lines are darker, because they are more accessible. Think of
this plot as showing how many non-mandatory destinations can I get to by transit from my zone
compared to some other zone. Transit accessibility to non-mandatory destinations often shows
up in auto ownership models. Households that are in darker areas tend to own fewer cars because
their transit accessibility is much better than lighter areas. So from a policy sensitivity
perspective, a model that has such a variable in it will predict changes in auto ownership due to
changes in transit accessibility.
496
Page 40
40
This plot shows transit access to non-mandatory destinations. This uses the walk and bike modes
as the measure of impedance, and non-mandatory purpose for the size term in the destination
choice model. Think of this plot as showing how many non-mandatory destinations can I get to
by walking and/or biking from my zone compared to some other zone. More compact, dense
communities with close proximity to retail and service employment show up darker on this map.
Non-motorized accessibility to non-mandatory destination variables also often shows up in auto
ownership models. Households that are in darker areas tend to own fewer cars because their nonmotorized accessibility is much better than lighter areas. Again, a model with such variables in it
will show sensitivity to mixed-use and density-oriented land-use policies.
497
Page 41
41
This plot shows access to non-mandatory destinations specifically for 0-auto households. This
variables takes into account all modes of travel auto, walk\bike, and transit. However, the
modes are based upon the mode choice model for someone without access to their own
automobile. Therefore, non-motorized and transit modes influences the accessibility much more
than auto (though auto still influences the accessibility due to ride-sharing with non-household
members). Sometimes auto ownership models use a number-of-autos specific destination choice
log-sum instead of different mode-specific log-sums. The drawback to the use of a autoownership-specific log-sum is that the calibration of the mode choice model can affect the logsum (since alternative-specific constants for each mode are changing). And, changes in the logsum can affect auto ownership, which affects mode choice shares. Therefore the calibration
process must be iterative and can be time-consuming. For this reason often mode-specific logsums are used as explanatory variables in upper-level models such as auto ownership and tour
generation.
498
Page 42
499
Page 43
43
OK, so far weve talked about the influence of mode and traveler characteristics on
accessibilities. Now lets talk about differences in time-of-day. It is clear that one of the key
advantages of an activity-based model is consideration of multiple time periods in travel
decisions. For example, levels of congestion experienced in the PM peak, on my way home from
work, can and should influence when I leave for work in the morning. Many aspects of the
transportation network vary throughout the day, including the modes available, the levels of
congestion, the availability and service frequency of transit, whether tolls are assessed and their
value, the land-uses available, perceptions of safety, reliability, and so forth.
In addition, land-usage varies by time-of-day. Stores, doctors offices, and other business
establishments are only open during certain times of day. We will now explore how
accessibilities can be calculated throughout the day.
500
Page 44
Destination
Choice
Re-choose time-of-day
based upon chosen
destination and socioeconomic variables
44
One method of account for time-of-day effects in daily accessibilities is to re-simulate the
choice. This is akin to the feedback process used in models where skims are fed back to demand
models and the models are re-run until the travel times output from the models are consistent
with those input to the models. In the example shown, a simplified time-of-day choice model is
applied in which the destination of the tour is unknown. The time periods assigned from this
simple model are used to calculate the probability of choosing a destination. Once the destination
is chosen, the time-of-day choice model is re-run to choose the final time period. The
disadvantage of this approach is that the initial choice of time-of-day will be inconsistent from
the full time-of-day choice for a number of observations, and it is unclear what effects this might
have on model outcomes.
501
Page 45
Period (p) =
AM Outbound/
PM Return
Periods (p) =
Midday Outbound/
PM Return
Periods (p) =
PM Outbound/
Evening Return
Logit Average: Mode choice logsums are averaged (with constants) across several
(or all) time period combinations p, for zones z and modes i
Z P I U z , p ,i
ln e
1
p
1
i
45
A logit average of all relevant time-of-day period combinations can be taken, in order to ensure
proper representation of the effects of network level-of-service across the entire day in a daily
average accessibility. This is the mathematically correct way to represent multiple lower level
choices in an upper level logit model, but also imposes the most computational burden.
502
Page 46
Time-space Prism
Time-space prism
One way to think about accessibility is to consider the Time-Space prism that Geographers
defined 40 years ago. This sketch shows a driver in space (horizontal axis) and time (vertical
axis). The prism shows the amount of time available between departure from an origin and return
to the same origin. The orange color line is the one dimensional space a person can reach within
the available time and this is named the potential path area. If we add some more realism to this
representation we get the next figure.
503
Page 47
Activity Scheduling
47
This time we added the activity schedule of a person and represent space in two dimensions. We
also added two major anchor points in the schedule that are home and work.
The shaded areas around home and work are potential path areas centered at anchor points. This
is the type of accessibility that we compute around destinations. For example, around work
assuming we have 15 minutes available time our accessibility will be the amount of stores we
can reach and/or the amount of restaurants or any other type of activity locations.
This figure also shows another type of accessibility (the blue shaded areas). We show here an
area that can be reached during the morning commute and during the evening commute. These
look more like a corridor accessibility in essence an area of the urban environment that can be
reached on the way to work and another area that can be reached on the way back from work to
home. We call these time-space based, and they depend heavily on the available time from one
place to another and on the activity schedule of an individual.
504
Note that there are many places which a person cannot reach within the available time and even
when the person can reach them there may be no activity location available (these areas are
called urban desserts)
In this webinar we will discuss both types of accessibility - anchor-based and time-space based.
505
Page 48
AM peak
Midday
PM peak
Night time
50.00%
40.00%
30.00%
20.00%
10.00%
0.00%
48
So far we looked at the impact on accessibility of travel time (and speed on the network). But,
we also know there is another temporal aspect that we need to account for time of day. There
are dramatic differences in the amount of opportunities a person has for activity participation at
different times of a day. Just think of post offices at 6:00 am and at 11:00 am.
One way to show this in accessibility indicators is to label every stores by opening and closing
hours and then do some post-processing. Unfortunately this information is not readily available.
Another option is to se information from a travel survey that also asks the industry in which
every person works using standardized codes of industries. Then, based on weighted data (to
represent the population) of the survey develop time of day profiles of percent of people present
at work at different times of a day. This can represent the amount of opportunity available to
people to pursue an activity.
The next few slides show hour by hour the number of persons employed in retail in one portion
for the Southern California region.
506
507
Page 49
1am
Beverly Hills
Santa Monica
49
We show the location of Santa Monica at the Pacific Ocean, further northeast of Santa Monica is
Beverly Hills and further east is downtown LA. This also happens to be the most popular and
crowded corridor in the greater Los Angeles area.
Areas in blue have no opportunities; light blue have a few; yellow indicates more opportunities;
the orange color shows there are even more; and red is used for a very high-opportunity location.
508
Page 50
1am
50
At 1 am there are still some people working in retail along major freeways.
509
Page 51
2am
51
At 2 am most people have gone home, with just a few pockets of retail still available.
510
Page 52
3am
52
511
Page 53
4am
53
512
Page 54
5am
54
More and more people arrive at retail locations. For this reason, there is an increase in the
number of person in retail per square mile.
513
Page 55
6am
55
514
Page 56
7am
56
515
Page 57
8am
57
By 8 am, the majority of the areas seems to have reached a high number of retail employees.
Note the corridor from Santa Monica to Beverly Hills and then Downtown LA has high numbers,
and a few blocks show very high peaks in red.
516
Page 58
9am
58
Similarly at 9 am
517
Page 59
10am
59
518
Page 60
11am
60
519
Page 61
12noon
61
520
Page 62
1pm
62
521
Page 63
2pm
63
522
Page 64
3pm
64
523
Page 65
4pm
65
524
Page 66
5pm
66
525
Page 67
6pm
67
526
Page 68
7pm
68
527
Page 69
8pm
69
528
Page 70
9pm
70
529
Page 71
10pm
71
530
Page 72
11pm
72
531
Page 73
12midnight
73
532
Page 74
8PM
74
In this slide we see a clear difference between the presence of retail employee at 11 AM and at 8
pm. These differences in accessibility at different times during a day can be used as explanatory
variables in activity and travel participation model components of activity-based travel models.
533
Page 75
80.00%
70.00%
60.00%
50.00%
40.00%
30.00%
20.00%
10.00%
0.00%
3
Percent used to
compute Min
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
Time of Day
Most traffic assignment four step models either provide peak period travel times that for our
purposes contain extreme differences in available opportunities as we can see in this image.
During the AM peak period we could be reaching 10% or 50% of the employees in this specific
industry. Nevertheless, the four step model provides just one set of travel times on the roadways.
For this reason we compute the minimum and maximum within this period in an attempt to
capture this variation.
In other applications, and when hour-by-hour travel times are made available instead of the minmax computation, we can calculate accessibility at each hour of the day.
534
Page 76
AM Peak Max
Finance
Midday
Max Finance
PM Peak Max
Finance
Night Time
Max Finance
76
Combining the availability of opportunities in space and time that we reviewed before with the
information about travel time on the network, we obtain similar accessibility indicators for the
number of employees of an industry that can be reached with ten minutes from each origin point
(in this example the block centroid). This slide shows the accessibility within ten minutes to
finance employees at different periods in a day. As expected at night time (after 7:00 pm) very
few finance opportunities are available, although the travel times are shorter and the network is
not as congested as during the pm peak period.
535
Page 77
AM Peak Max
Retail
Midday
Max Retail
PM Peak Max
Retail
Night Time
Max Retail
77
Similarly on this slide we see the accessibility by time of day of the retail employee. This time,
however, retail is far higher than finance in the late evening hours. In activity models we should
expect this time of day pattern when we model activity types.
536
Page 78
78
It is also interesting to map accessibility in correspondence of major freeways. This slide shows
the clear impact on accessibility of intersecting freeways in the evening. These are the locations
of gas stations and convenience stores.
537
Page 79
79
Just superimposing transit routes over the accessibility maps we can see clearly patterns of
multiple routes over higher-density environments.
Computation of accessibility using transit information can be done using more detail. We have
done this using time tables of transit agencies in Southern California. The key difference from
roadway accessibility is that in transit we need to account for walking time to and from bus stops
and to develop an algorithm that accounts for vehicles headways and overlapping routes. You
can find more details in a TRB paper with first author Ting Lei.
538
Page 80
80
Travel by transit is bi-modal in nature. Both the transit network and the pedestrian network
define possible routes of a transit trip. Therefore, the first step for modeling transit travel is to
develop an integrated network containing both transit and regular road links. In practice, the
network for regular roads are often readily available in a Geographic Information System. The
transit network (with schedules) seldom exists in GIS. Instead, they are often provided in tabular
format in terms of the schedules and coordinates of transit stops or timed points.
The first task to make the analysis work is to build a transit network from tabular data. It is
typical in practice for transit authorities to publish schedules only for a subset of points of any
given route called its timed points. They may or may not be actual stops. Therefore, the
schedules for the rest of the stops in a route need to be inferred from the schedule-building
program, e.g. based on their (linear referencing) distances to the immediate neighboring timed
points. Once the schedules are generated, they are aggregated and stored as attributes of transit
links.
539
The next step is to merge transit network with the regular road network. This is required because
the positions of transit stops from the transit database will not typically match the road network
since they come from different sources. Discrepancies between the two networks are shown in
the figure below; and they are removed using GIS functions.
540
Page 81
81
The combined bi-modal network thus obtained contains routes for all transit modes in the
LACMTA database and it consists of 1,748 routes and 89,980 stops.
541
Page 82
LACMTA Trip
Master Database
(Q4 2008)
Attributes
Build transit
schedules from
trips file
Links
Transit Network
Schedule-based
shortest path (ArcGIS
module using C)
Block centroid
shape files:
centroids attached
to nearest network
nodes
Counting
opportunities
within pre-specified
time standards
82
The overall workflow for computing transit accessibility is shown here. After the bimodal
network is constructed, connectivity (topology) relations among transit and road links are
compiled using GIS software. In this application, we used ESRIs ArcGIS. This makes the
bimodal network a graph in the sense of graph theory, except that our network contains
schedules in addition to the lengths of arcs in regular graphs.
A key component of this research is a shortest time path algorithm that can take advantage of the
schedule information stored in the augmented network. Considering the large number of origins
and destinations and the need to compute routes for all pairs of them, we implemented the
schedule-based shortest path algorithm reported in a previous paper (Lei and Church, 2010) by
means of adapting an efficient open source C++ library for graph algorithms.
The basic procedure is based on augmenting the well-known Dijkstra algorithm as follows:
1. Set the label for the source node to zero and the labels for all other nodes to infinity. Mark all
nodes as unvisited.
542
543
Page 83
Data Source: LACMTA master trip database, Ting Lei, Jan. 2012
83
This isochronic map shows the space-time prism of travel by transit, starting from a given
location. The distorted shape of the space-time prism footprint which follows transit routes and
stops distinguishes travel by transit from travel by other modes, such as the automobile.
544
Page 84
Midday
PM Peak
Night Time
Maximum number of reachable education employees for a 20-minute buffer by transit by time of day in Central Los Angeles.
Maximum number of reachable education employees for a 20minute buffer by time of day in Central Los Angeles by Car
84
The figures in this slide show transit accessibility indicators for the retail and education type and
20 minute time buffers for four periods in a day.
For comparison, we also show automobile accessibility indicators computed for the same region.
From these figures, we can see the difference of accessibility for different types of opportunity,
different times of day, and different modes of transportation.
545
Page 85
85
Fine resolution accessibility indicators can also be used in activity-based models, via parcelbased data. This plot shows number of employees reached within 10 minutes by each parcel.
Such information can be useful to measure the attractiveness of parcels for residential location
choice models.
546
Page 86
86
Fine resolution accessibility indicators can also be used to explain in a visually clear manner why
we want land use policies that promote higher density and mix of activities and why cul-de-sac
type of developments provide low levels of accessibility and motivate people to use their car to
reach locations of greater opportunity. This slide shows the difference in accessibility enjoyed by
different urban forms. The small parcels coupled with a grid street system in the middle of the
map lead to better accessibility than the larger parcels located in the cul-de-sac neighborhood in
the lower right portion of the map.
547
Page 87
Network \ Modes
Different accessibility measures are based upon certain modes or taken
across all modes, depending on purpose of accessibility measure
Mode choice logsum used when accounting for all modes
Time-of-day periods
Period-specific or logsum across periods
Activity Purposes
Related to activities available
Land-use variables used in destination choice size terms
Activity types considered in other accessibility measures
Spatial Unit
Zones, micro-zones, or parcels
Activity-Based Modeling: Accessibilities & Treatment of
Space
87
Now weve discussed all of the major elements of accessibilities. How do we put it all together
to calculate accessibility measures? One way is to define the various combinations of each
element. This includes:
Socio-economic markets: The socio-economic category is important to know the value-of-time
(and therefore determine the cost coefficients to apply to tolls, transit fares and other out-ofpocket costs), and also can be used to determine alternative-specific constants for modes.
Network\mode: Accessibilities can be mode-specific. They can be calculated taking into
account groups of modes (such as all transit line-haul modes), or they can be calculated across all
modes.
Time-of-day periods: Accessibilities can be calculated for specific times of day, such as 9:30
A.M., or calculated across multiple time periods, reflecting a daily accessibility.
Activity purposes: Accessibilities are typically based upon travel for a specific activity purpose,
such as shopping or eating out. The activity purpose relates to the size term that is considered in
548
the destination choice model. An accessibility for shopping would typically be influenced by
retail employment, for example.
Spatial unit: The treatment of space is an important consideration, and often imposes the most
computational burden.
For example, a parcel-based approach to calculating accessibilities imposes substantial
computational overhead, as accessibilities would need to be calculated for parcels2 cells, times
the number of markets, modes, time-of-day, and activity purposes. Assuming 1 million parcels
and time in 30 minute increments, this might amount to more than 1 trillion calculations! For
this reason, often trade-offs have to be made in terms of the representation of space and the
temporal dimension when constructing accessibilities.
549
Page 88
Description
Travel Cost
Access to nonmandatory
attractions by SOV in off-peak
Car
ownership
Total weighted
employment for
all purposes
Access to nonmandatory
attractions by transit in
off peak
Car
ownership
Total weighted
employment for
all purposes
49
Access to nonmandatory
attractions by all modes
Daily Activity
Pattern Model
Total weighted
employment for
all purposes
1618
Access to eatingout
attractions by all modes
except SOV
Joint tour
frequency
Weighted
employment for
eating out
43-44
Individual
subtour
frequency
Weighted
employment for
at work
88
Here is an example of how the accessibilities for the SANDAG model were defined. There are a
total of 50 different origin-based accessibility measures. They are defined based upon the period
of travel considered, the modes considered, the market segmentation considered, and the activity
purposes considered. The spatial system used is the micro-zone. Software is optimized so that
various elements of accessibilities, such as the travel cost variables or the size term variables, are
re-used to the maximum extent possible. This results in accessibilities that can be calculated in
less than 10 minutes (for 50 variables, across 22,000 micro-zones).
550
Page 89
Land-Use Data
Zone, micro-zone, and/or parcel level
Households, population, employment by type
4D measures intersection density, walkability, mixed-use, etc.
Future forecasting
Hold coefficients constant, vary input networks and land-use data to
measure effects of accessibilities on travel
Activity-Based Modeling: Accessibilities & Treatment of
Space
89
Here are the data needs required for calculating accessibility variables. Networks are required by
time-of-day. They are skimmed for each time period of interest. Since various components of
time and cost are saved, this can result in a large number of matrices.
Land-use data is required at the zone, micro-zone, and/or parcel level. This typically includes
households and population by type (which must be consistent with the synthetic population) and
employment by type. Land-use design and density information is also required.
Household survey data (and transit on-board survey data) is used to determine the coefficients on
in-vehicle and out-of-vehicle time and cost, socio-demographic variables, and size term
coefficients which state the relative attractiveness of different types of land-use for different
activity purposes. Household and/or transit on-board data is also used to estimate the effect of
accessibility variables on travel behavior. For example, how sensitive is auto ownership to retail
employment accessibility? How sensitive is the number of tours generated to the accessibility of
home and work? We will explore these relationships further in subsequent webinars.
551
In forecasting, the coefficients are typically held constant, and the input networks and land-use
data are changed to reflect forecasted network supply and development patterns (either assumed
land-use or predicted from a land-use model). The model system is then run to determine the
effect of accessibilities and other changes in the model inputs on travel behavior and network
performance.
552
Page 90
90
This slide shows different types of accessibilities and what model components they effect in the
Sacramento, California version of the DaySim model. Besides the standard level-of-service
matrices (times and costs) that are used in mode and destination choice models, the system relies
on two key types of accessibility measures: mode choice log-sums and destination (a.k.a. modedestination) choice log-sums. The mode choice log-sums vary by whether they are for sampled
or chosen destinations, and what purposes and time period(s) are considered, whereas the
destination choice log-sums vary by purpose, typically with weighting across periods.
553
Page 91
91
This slide shows the same sorts of accessibilities as the previous slide, and their consideration in
the San Diego version of CT-RAMP. There are some differences between the exact purposes and
time period definitions between the model systems, but also a good deal of consistency in their
general definition and impact on the various model components.
554
Page 92
Feedback
Primary effects:
Route and mode
choice
Secondary effects:
Time-of-day and
destination choice
Tertiary effects:
Tour\stop generation
and mobility models
92
Here is a practical, concrete example of how an activity-based model might respond to network
scenario. In this scenario, we are modeling increased transit service in a congested corridor
during the AM and PM peak periods. The change is coded into our transit network and the model
system is run. The first change one might see is that level-of-service skims for affected zonepairs will reflect the improved transit in-vehicle and out-of-vehicle times. As a result of these
network (and skim) improvements, transit has a higher probability of selection in route choice
(transit riders with a choice of routes will choose the improved transit routes more) and in mode
choice (transit will be chosen more frequently for trips and tours in the corridor). These are all
primary effects of the transit improvement, and these effects are typically considered in a fourstep trip-based model.
An activity-based model will also typically consider secondary and tertiary effects, that the fourstep model does not. These include time-of-day and destination choice impacts. The transit
improvement will produce a higher-value mode choice log-sum for affected zone pairs. The
time-of-day choice model considers the improvement in mode choice log-sum, causing more
555
transit riders to choose to travel in the AM and PM peak periods (as a result of the improvement
in log-sums). With feedback, some auto travelers may shift back into the peak periods since
some of them have chosen to switch to transit, freeing up some capacity. These effects are
represented by the feedback loop to the left of the diagram. The improvement in the mode choice
log-sum also affects destination choice; more travelers will choose destinations in the improved
corridor, since the relative accessibility of this corridor has improved.
The improvement in the destination choice log-sum has tertiary effects on tour and stop
generation models and medium-term mobility models. Households that reside in the corridor
may generate more direct tours with less stops per tour, as they change their travel patterns to
take advantage of the transit service. Households may opt to own fewer cars and more transit
passes as a result of the improvement. All of these potential travel behavior changes are
represented in activity-based models with well-formulated accessibility variables.
556
Page 93
Learning summary
93
In todays session, we covered the treatment of space and the calculation of accessibilities in
activity-based models.
Accessibilities variables are important because they convert the times and costs provided by the
transportation network to the land-use system. They are used to represent the alternatives
available in lower level model components, such as mode choice, on upper-level model
decisions, such as the number and complexity of tours generated and the number of autos owned.
The important dimensions of accessibilities include the level of detail in the spatial system, the
networks and modes considered, the time periods for which accessibilities are calculated, and the
activity purposes and/or land-uses considered.
The main types of accessibilities used in activity-based models are buffer-type variables, mode
choice log-sums, and destination choice log-sums.
557
Page 94
558
Page 95
February 2
February 23
March 15
April 5
April 26
May 16
June 7
June 28
July 19
August 9
August 30
September 20
95
Once again, here is the schedule for the webinar series. Our next webinar, three weeks from
today, will cover long-term choice models such as work and school location choice, and
medium-term mobility models such as auto ownership, transit pass-holding, and parking
reimbursement.
Thank you!
559
Page 96
96
560
not developed based upon workplace survey data, but such data may be useful to validate the
predictions of workers at specific workplaces or for different types of employment. There are
other models that do use occupation type when choosing workplace. Workplace location choice
will be covered in detail at the next webinar in 3 weeks.
562
563
Page 1
Activity-Based Modeling
Session 7: Long-Term & Mobility Choice Models
June 7, 2012
564
Page 2
Acknowledgments
This presentation was prepared through the collaborative efforts of Resource
Systems Group, Inc. and Parsons Brinckerhoff.
Presenters
Maren Outwater and Peter Vovsha
Moderator
Stephen Lawe
Media Production
Bhargava Sana
Resource Systems Group and Parsons Brinckerhoff have developed these webinars
collaboratively, and we will be presenting each webinar together. Here is a list of the persons
involved in producing todays session.
Maren Outwater and Peter Vovsha are co-presenters. They were also primarily
responsible for preparing the material presented in this session.
Stephen Lawe is the session moderator.
Content development was also provided by Nazneen Ferdous, John Gliebe and Joel
Freedman. John Bowman provided some content and served as a reviewer.
Bhargava Sana was responsible for media production, including setting up and managing
the webinar presentation.
565
566
Page 3
February 2
February 23
March 15
April 5
April 26
May 16
June 7
June 28
July 19
August 9
August 30
September 20
3
For your reference, here is a list of all of the webinars topics and dates that have been planned.
As you can see, we will be presenting a different webinar every three weeks. Three weeks ago,
we covered the sixth topic in the seriesAccessibility and Treatment of Space.
Todays session is the fourth of nine technical webinars, where we will cover the details of long
term and mobility choice models. In todays session, we will prepare participants for developing
long term and mobility choice models and discuss which models are important for different types
of planning studies. In three weeks, we will cover activity pattern generation models.
567
Page 4
Learning Outcomes
By the end of this session, you will be able to:
Determine which long-term and mobility decisions are
important to model
Describe how long- and mobility decision models are integrated
into the activity-based model system
Consider the benefits, costs and key challenges of modeling
long-term and mobility decisions
This session covers a long term and mobility models for activity-based modeling systems. Today
we will be determining which long and medium-term decision models are important for different
types of policy studies and for different types of transportation alternatives. We will also
describe how long term and mobility decision models are integrated into the larger activity-based
modeling system so that you can understand the context of how these models work with other
modeling components. The long and medium-term decision models are at the top of the
hierarchy of the activity-based modeling system and incorporate feedback from several of the
downstream model components. Another important topic for today is the benefits, costs, and key
challenges of modeling long and medium-term decisions. Most activity based models in use
today include a limited set of long and medium-term choice models but many of those under
development are including an expanded list to provide greater sensitivity for the travel demand
models.
568
Page 5
Session Outline
In this session we will cover
Why modeling long-term and mobility decisions is useful
Where long-term and mobility decision models fit into travel
model systems
How these models are developed
What data sources are needed
The benefits and costs of system integration
In this session, so we will cover why modeling long and medium-term decisions is useful to
provide added sensitivity in the models to policies and transportation investments under
consideration. We will present information on where long and medium-term decision models fit
into travel modeling systems and describe how feedback is used to represent accessibility
measures and travel times and costs for these long and medium-term choice models. We will also
cover how these models are developed in practice as well as several new areas of research that
are being considered for these models. We will touch on what data sources are needed for model
estimation, application, and validation. It is useful to note that most of the data sources required
by these models are provided by the synthesized population and by the accessibility and log
some measures discussed in the previous two webinars. Lastly, we will present some of the
advantages and disadvantages of network integration within the context of the activity-based
modeling system.
569
Page 6
Terminology
Before we begin, I would like to cover a few terms that we will be using in this webinar.
Usual workplace location choice models are used to determine the workplace that you
visit most often during a week. A separate process can be used to identify the workplace
that you visit on a given travel day.
Usual school or college location choice models are used to determine the location of the
school or college that you visit most often during a week. Again, a separate process may
be used to identify the specific school location you visit on a given travel day if it is
different from the usual location.
Vehicle availability models describe the number of vehicles available to a household. In
this context, vehicles may include passenger cars, SUVs, pickup trucks, and motorcycles.
The term availability is used here to identify any vehicles that are owned, leased or
borrowed on a long-term basis and readily available in working order for travel by
members of the household. This is a departure from the prior use of the term auto
570
571
Page 7
OK,
thattosounds
I have
pick up
great.
Ill
Emma fromtake
school
light
work
andrail
goto
grocery
today.
Bye honey,
shopping
later; I
see
youthe
later.
need
car.
Long and medium-term mobility models are intended to capture choices that a household or
members of the household are making regarding longer-term decisions such as buying a car and
medium-term decisions such as whether to buy a transit pass or a toll transponder. These
decisions are longer-term because they affect travel choices over the course of the year, as well
as influencing travel choices on a day-to-day basis. This illustration shows a three person family
with one car considering trade-offs between individual activities and what modes can
accommodate these activities. Each person in the family has different needs at different times, so
they sort out who will use the car at which times and for which purposes.
572
Page 8
Why are these long-term and medium-term mobility choices important? Primarily, it is because
they directly influence other travel choices like mode and activity scheduling. More recent
activity based model development has recognized the usefulness of these variables in the daily
activity pattern and mode choice modeling components. In addition, we recognize the usefulness
of providing a model to estimate the probability of these choices, so that we can represent the
influence of various policies and programs affecting these choices. For example, travel demand
management strategies, such as flex time and telecommuting, provided by many employers will
influence workers activity schedules and mode choice. In addition, pricing strategies will impact
workers differently if employers are subsidizing aspects of their travel. Usual work and school
location choice models reflect long-term decisions for workers and students that may be different
than the short-term decisions made about going to a meeting, or a seminar at a different location
on a given day. In addition, these models recognize the growing trends to work at home or take
classes online, either on a permanent or occasional basis. Again, these decisions are made in a
long-term context or within the travel model on a day-to-day basis to recognize that different
factors influence these decisions differently. The use of these mobility choices within the activity
573
generation, scheduling, and mode choice models has improved the explanatory nature of these
travel demand elements. Lastly, these models provide an ability to forecast changes in these
variables as a result of transportation investments and policies in the future.
574
Page 9
After developing these long-term and medium-term mobility choices we are able to test different
scenarios, such as travel demand management program participation and demonstrate the
changes in performance of the system. Many decision makers are interested in understanding the
influence of toll transponders, transit passes, and fuel efficient vehicles on travel demand and the
environment. These models allow that sensitivity.
One challenge in developing these models is that many of these variables are not wellrepresented in older household surveys, and so other data sources may be needed to develop
these models. The good news is that many newer household surveys are including questions
about these mobility choices, so that they can be readily included in the activity-based modeling
system. An example would be asking questions about whether someone has an alternative work
schedule (like working four 10-hour days) or what their usual mode to work is. Some areas have
surveys that were designed specifically to evaluate a particular program like the travel demand
management program and so may be useful in developing some of the mobility choice models.
575
Since many of these models are quite new, we dont yet understand whether parameters are
transferable or whether behavior is similar in different areas.
576
Page 10
Build Alternative(s)
10
Lets consider a transportation planning and policy project that might be faced by an MPO or
DOT and how long-term decision making and mobility fits into the picture. We used this
example in the population synthesis webinar and will discuss how future topics will be affected
by this bridge expansion example.
For this scenario analysis, we will be considering a number of alternatives: a no-build alternative
and a various configurations of the build alternative. In the no-build alternative the bridge has 4
lanes (2 in each direction), there are no tolls, and the transit fare stays the same all day. In the
various build alternatives, there are 6 lanes on the bridge. In some alternatives the two additional
lanes will be HOV lanes all day, while in other alternatives the two additional lanes will be HOV
lanes only during peak periods. In addition, in some build alternatives there will be a new toll
that is the same across the entire day, while in other build alternatives there will be a toll that will
be only applied during peak periods, or when certain levels of congestion occur. Finally, in the
build alternatives regional transit fares will be higher during peak periods.
577
Page 11
11
The bridge expansion example will have varying effects on the long-term and medium-term
mobility choice models. For example, vehicle ownership may decrease with tolls, and vehicle
type choice may increase if fuel efficient cars are offered discounted tolls. If someone owns a
transit pass they may increase their transit use across the bridge to take advantage of the carpool
lanes and the free passage. If someone owns toll transponder, they may be more likely to cross
the bridge on a regular basis. It is also possible that work locations may change over time as
people decide that paying a toll five days a week to get to work doesnt fit within their
transportation or household budget. Less likely, but still possible, school locations may also
change if they are across the bridge and a toll is required. Work and school locations are also
affected by tolls that are higher in the peak period because workers and students are often
required to travel at this time.
578
Page 12
12
There are likely going to be other indirect effects of the new toll bridge, if traffic volumes are
spread throughout the day and travel times improve as new workers can consider work locations
across the bridge. Other workers may consider telecommuting as a viable option, rather than pay
the toll. These shifts will be different for higher and lower income families and those who have
work schedule flexibility to travel outside the peak period.
School location may be less impacted by a bridge expansion example, but college locations may
be influenced in much the same way as work locations. Private schools, though, represent an
element of school location that may be influenced by a bridge expansion.
579
Page 13
Vehicle choice
Hybrid and other green vehicles may incur a smaller toll (thus,
generate less revenue) relative to regular/less fuel efficient vehicles
Fuel efficient vehicles may be chosen for bridge crossings more
often than other cars, especially in the peak when tolls are higher
Other vehicles may be chosen for bridge crossing trips during offpeak periods when tolls are lower
Activity-Based Modeling: Long -Term & Mobility Choice Models
13
The bridge expansion example may also affect vehicle choices in the household. As the price for
travel increases on the toll bridge, lower income households may choose to reduce car ownership
in favor of alternative modes such as transit. Higher income households may appreciate the
improved travel times on the bridge and over time purchase additional vehicles. Households with
more people may choose to take advantage of carpool options.
Some toll bridges are considering offering discounted tolls for hybrid or other green vehicles,
which may increase the use of these vehicles as well as the purchase of these vehicles. People
may also choose to use fuel efficient vehicles during peak periods when tolls are higher and to
use other vehicles during off-peak periods when tolls are lower.
580
Page 14
14
The bridge expansion example also affects some aspects of personal mobility directly and other
aspects indirectly. For example, transponder ownership may increase significantly if
transponders offer a discount for frequent use. Travelers who are considering transit as an option,
may be more likely to purchase a transit pass to take advantage of the transit improvements on
the bridge. The bridge expansion example may have indirect and smaller impacts on drivers
licenses and bicycle ownership.
581
Page 15
15
There are also a series of worker mobility models that will be affected by improvements on the
bridge. For example, tolls that are lower during offpeak periods may encourage workers to
travel in off-peak periods by working longer hours or by shifting work hours. Workers may also
choose a work schedule that involves longer work hours for fewer days, such as a 10 hour, fourday workweek, or a nine-hour nine-day two-week work schedule.
Employers who have travel demand management programs may offer employees subsidies for
transit passes, parking, or other travel. These subsidies can influence travelers decisions by
reducing costs for specific alternatives. Some employers offer parking or travel subsidies as part
of an overall benefits package rather than to influence a mode shift. In the context of the bridge
expansion example, these subsidies can influence modal shifts especially if one mode is
subsidized and another is not.
582
Page 16
Person
Decision
Location
Models
Vehicle
Models
Personal
Mobility
Worker Decision
Student
Decision
Auto Allocation
Drivers License
Transit Pass
Worker
Mobility
Transit Pass
16
There are many types of mobility decisions in the activity-based modeling system made by
different agents. Household decisions include auto ownership, bike ownership, and toll
transponder ownership because these can all be used by any member of the household. Personal
decisions include which auto to use for each trip and whether to obtain a drivers license or a
transit pass. The auto allocation choice could be considered as a medium-term decision, if the
auto is driven primarily by one person most of the time. If the auto is shared among two people,
then the auto allocation choice becomes a shorter term decision where the auto may be used by
different people throughout the day. These types of shorter-term decisions are made within the
context of the daily activity and travel models rather than as part of the long-term and mediumterm mobility models.
Another example of this is our workplace and school or college location choice models where
the usual workplace and school or college locations may be chosen as part of the long-term
decisions and the workplace and school or college locations for a particular day may be chosen
as part of the daily activity and travel decisions. Here, we may have someone who works
583
downtown typically, but on the day of travel goes to a business meeting in another city. The
long-term workplace location downtown is modeled as a long-term choice, and the short-term
workplace location in the other city is modeled as part of the short-term daily activity pattern.
A third example of this is our work duration, work schedule, and usual mode models. These can
be included as both long-term mobility models and as short term activity travel models. Again,
the purpose of separating these choices is to identify the characteristics involved in a long-term
choices separately from the characteristics involved in the short term choices and to recognize
the differences between them.
The last set of mobility of models include those that involve travel demand management
programs such as flex time, telecommuting, and subsidies for transit passes, parking, and other
travel.
584
Page 17
17
The mobility models have a series of design decisions that will affect the model components,
model structure, and sensitivities in these models. The first part of my talk has focused on which
models, or choices should be represented in the modeling system. This part focuses more on
design decisions within each modeling component. These include identifying the choice to be
made or the variable of interest such as the locations of work activities. Most of the long-term
and medium-term mobility models are structured to produce a single variable for input to the
activity pattern and travel models, such as whether someone owns a transit pass. It is also
necessary to identify the variables that will impact this choice and these should include as many
of the policy sensitivities desired in the modeling system. An example of the variables that would
affect owning a transit pass are person type (workers, students), age, income, time and cost to
work. Another aspect of design is how it will fit within the modeling system. This could be as a
direct input, as a pre-condition, or as a combined decision with another variable.
585
Page 18
18
As I mentioned before, mobility decisions may be embedded within day pattern models and as
long-term choices, or represented in one or the other. This is true for work and school location
choice models. If the mobility decisions are included as long-term choices then questions about
usual patterns of travel to work and school locations are needed in the household travel survey to
distinguish these usual patterns from a daily pattern collected as part of the travel diary.
This is true also for work-at-home patterns representing people who work at home full-time, or
students who are enrolled in online colleges. Current household surveys capture people who stay
at home rather than go to work or school, but may not capture whether they work or attend
classes at home as a substitute for going to work or school.
Another more recent expansion of mobility models is in the area of usual work mode, duration
and schedule to distinguish this from the daily work mode, duration and schedule. For example, I
may take transit to work every day except on a day where I need to visit a client. So I have a
transit pass and take transit regularly, but on my day of travel I choose to take the car and go to
work as well as to visit the client and then return home. The reason I take my car is because the
586
client is not in a transit area but my regular workplace is in a transit area, so the factors for these
two destinations are different. We are still modeling an average day of travel, but can account for
those who deviate from their typical work schedule with this additional information.
587
Page 19
19
Some land-use models have components of mobility decisions built within them, such as
workplace or school location choice. There are several reasons one may choose to build these
mobility decisions into the land-use model instead of the travel model. First, this offers the
opportunity to formulate a joint residential and workplace location choice model, which has been
researched, but not put into practice. Second, a disaggregate land-use model can directly connect
each worker to each job and each student to each enrollment. One example of a workplace
location choice model like this is at PSRC in Seattle, where each worker chooses a job and
workers without jobs represents unemployment and there are also potentially jobs without
workers. In this example, the micro-simulation will match new workers to open jobs each year
and relocate some workers to available jobs. One benefit of this approach is that there cannot be
more workers than jobs at a location. Another area of research is in joint residential and vehicle
ownership choice models. The benefits of these models are not yet well understood, or whether
the additional complexity offers significantly better results. Home based jobs can also be
included in the land-use model to represent home-based businesses or services provided to
homes.
588
589
Page 20
Disadvantages
Assumes that households/persons of the same type will make similar choices in
the future
Does not account for important policy and cost variables that influence these
decisions
Cant do scenario analysis
Activity-Based Modeling: Long -Term & Mobility Choice Models
20
A third option for mobility decisions is to include them within the population synthesizer. These
might include auto ownership, drivers licenses, or people who work at home. These are most
likely to be included as uncontrolled population attributes, but could be controlled if desired. The
advantages of including these decisions in the population synthesizer are that it is easy and
doesnt require a separate model component. The disadvantage is that the population synthesizer
is not sensitive to policies or cost variables that influence these decisions and scenario analysis
will not be possible. As a result, most current activity-based models do not include mobility
decisions within population synthesizers.
590
Page 21
Vehicle ownership/availability
Common in both trip-based and activity-based models
Used as input to activity patterns, destination and mode
choice
Activity-Based Modeling: Long -Term & Mobility Choice Models
21
There are three mobility decisions that are included in the activity-based modeling system, at a
minimum, but these are quickly changing to include many additional decisions. The workplace
location choice model replaces the home-based work trip distribution in a trip-based modeling
system. The school and college location choice model replaces the home-based school and
college trip distribution in nature-based modeling system. The vehicle availability choice model
is common to both work based and activity-based models except that there are many additional
accessibility measures and log sums that may be possible in the activity-based model version.
591
Page 22
22
Workplace location choice models can be estimated, calibrated, and validated at different spatial
resolutions, including zones, micro-zones, or parcels. Workplace location choice models can
identify usual workplace location, day of work location, and work at home situations.
592
Page 23
Persons
Worker (FT/PT)
Occupation
Driver
Gender
Telecommuter
Land Use
Employment
Density by Type
Household
Density
Student
Enrollment
Mixed Use
Parking Density
Intersection
Density
Agglomeration
and Competition
Affects
Accessibility
Distance or
Distance Decay
Function
Mode Choice
Logsum
Mode/
destination
Logsum
23
Workplace location choice models may contain a variety of household, person, and land-use
characteristic variables as well as accessibility measures from the travel models. These variables
are also potential variables for school and college location choice models, as well as location
choice models for primary and intermediate stop locations within the activity-based model.
593
Page 24
Sales, marketing
Clerical, administrative, retail,
Production, construction, farming, transport
Professional, managerial, technical
Personal care or services
24
Workplace location choice models can be developed to segment workers and jobs in the same
occupation. This example from Phoenix and Tucson shows 5 occupation categories in the 2008
NHTS, which are correlated to the 26 categories of jobs by NAICS codes. In this way, you will
prevent salesmen from taking jobs in the manufacturing industry and someone in the
manufacturing industry from taking a retail job. This requires including occupation in the
synthetic population. Ideally, employment would also be classified by occupation rather than by
industry, since occupation and industry are weakly correlated.
594
Page 25
LN(D)
D0.5
D
D2
D3
50
40
Decay
60
30
20
10
0
0
10
20
30
40
50
60
70
50
60
70
Distance
2.5
2
Decay
1.5
1
0.5
0
0
10
20
30
40
Distance
25
Another potential variable in any of the location choice models is a distance decay function. This
reflects a non-linear relationship between distance and the location of interest (work or school).
595
Page 26
26
The impacts of the distance decay function vary based on the population. In this example, parttime workers look for local jobs; this tendency is most prominent in smaller regions like Tucson
and for shorter commutes (under 10 miles). Low-income workers look for local jobs and are less
specialized in occupation; the tendency is less prominent in small regions like PAG. Highincome workers do not look for local jobs; for MAG high-income workers could not be
distinguished from medium-income workers (baseline). These distance decay functions were
estimated from survey data in San Diego, Phoenix and Tucson.
596
Page 27
250,000
200,000
150,000
Normalized Estimated
Linear (trend)
100,000
50,000
50000
100000
150000
200000
250000
27
The results of a typical work location choice model may look similar to those from a trip-based
model for work trips. These include work trip length and distance distributions and district to
district flows. The work-at-home element is typically not included in a trip-based model but is
useful to calibrate in the mobility models. The example on the left shows a typical validation of
the work trip length distribution in Fresno and the example on the right is a district to district
chart.
597
Page 28
Work at Home
Work at home is rapidly growing because of
Communications technology
Structural shifts in occupation and industries
28
598
Page 29
29
The school and college location choice models follow many of the same features as the work
location choice models in terms of spatial resolution, formulation, and integration with the
activity-based model. Some models use actual school locations at parcels, but the advanced
practice to match students with enrollments directly has not been done in practice. This is
primarily because current practice of using multinomial logit models has worked well and
because the student-enrollment market is smaller than the worker-job market.
Some of the more recent methods to include long-term location choice models for usual schools
or colleges along with short-term choice models for daily travel to school and colleges are just
coming into practice. For example, Sacramento, Seattle, and Philadelphia are adopting this
approach in their models under development.
Many existing models separate school locations by grade level and to elementary schools, middle
schools, high schools, colleges, and some even include preschools. In the public school setting,
models can adopt a deterministic approach to send kids to the closest school. This often doesnt
work for private schools or colleges, where a multinomial logit choice model is more practical.
599
The choice between these methods should be determined by the student requirements for each
school (i.e. are students required to live in a district?).
600
Page 30
1 car
Households
2 car
3+ cars
Multinomial Logit
With No Vehicles
1 Car
Nested Logit
With Vehicles
2 cars
Households
3+ cars
Ordered Response
With No Vehicles
Households
1 car
With Vehicles
2 cars
2+ cars
3+ cars
Theoretically appealing
May be more difficult
to estimate
30
Vehicle availability or auto ownership models are typically developed in a multinomial logit
form, but other forms have been tested and researched. The most common multinomial logit
form has four alternatives beginning with zero car households.
A nested logit model could be used to separate households with and without vehicles and then
estimate the number of vehicles for those households with vehicles, but this is more difficult to
estimate and less theoretically appealing than an ordered response.
A third alternative form is an ordered response model where households are separated into those
with and without vehicles initially and then separated into one car and more than one car
households and finally separated into two car or three or more car households. Ordered response
models have been used in practice, but are more difficult to estimate than multinomial logit
models. Most agencies can develop multinomial logit models for vehicle availability, unless
there are issues with estimation or calibration using this process, whereby another approach can
be considered.
601
602
Page 31
603
Page 32
32
I have focused this discussion mostly on the commonly used mobility choice models, but there
are many additional models under consideration were under development at several MPOs
around the country. The most important mobility models for mode choice are driver license
holding models, transit pass holding models, bike ownership or availability models, toll
transponder models or limited-mobility models. The travel demand management program
participation models are quite useful for input to daily activity pattern models as they affect
scheduling for work trips directly. Some of the travel demand management program models are
also quite useful in mode choice. Another category of models under consideration or under
development are vehicle purchase or usage models that predict the type of vehicle a household
made by and which vehicle is used by which person in the household for which trip. These
models support the estimation of emissions and can also influence mode choice. In addition,
there are models to estimate the usual mode to work, along with the arrival and departure times
to work, which inform the actual mode and schedule for a worker on the day of travel. These
may change, of course, if the work location on the day of travel changes or if other locations on
the day of travel are not typical, like if I need to go to the doctors on the way home.
604
Page 33
33
The purpose of these additional mobility models is to predict attributes of mobility that help
explain the behavior of other travel choices and to provide sensitivity to specific policies of
interest. These policies are often employer policies, so are naturally focused on commute travel.
One popular example are commute trip reduction programs, which can include transit passes,
flex time, carpool parking spaces, and other subsidies to discourage travel in single occupant
vehicles during the congestion peak periods of travel.
605
Page 34
Level of participation
Flexible hours
Telecommuting
Carpool/vanpool
Parking subsidies
Subsidized transit passes
Different incentive
structures
Activity-Based Modeling: Long -Term & Mobility Choice Models
Travel demand management programs have been in place for many years in some cities, but are
often not well represented in trip-based models and included as a long-term choices in some
activity-based models. Current long-term choice models can represent different types of
programs and different incentive structures, but may not adequately represent the level of
participation, either by employers or employees. There is still much work to be done.
606
Page 35
Parametric:
Discrete (Logit) Choice Models (MNL, NL, CNL) applied for
each mobility attribute
Joint Choice (Logit) Models of several mobility attributes
with trade-offs
Multiple Discrete-Continuous Extreme Value (MDCEV)
Models (generalizations of logit) where choice of mobility
attributes is combined with some measure of use (VMT)
Activity-Based Modeling: Long -Term & Mobility Choice Models
35
There are quite a few different models that have been developed for modeling long-term and
mobility choices, although multinomial logit choice models is still the most commonly used
method. We will describe each type and provide some examples of mobility models.
607
Page 36
Drawbacks
36
The second half of this webinar will focus more on the methods used in modeling long-term and
mobility choices and on providing a few examples of these models in practice.
There are several nonparametric modeling methods, including population synthesis, which I
mentioned earlier, or assuming future distributions are similar to today. Both of these lack
sensitivity to transportation investments or policies. Some sensitivity may be incorporated
through the use of cross classification models, but the dimensionality will be limited to the
number of variables that can be included. These methods are very easy to implement and fast to
run, but do not take advantage of the disaggregate data or accessibility measures available within
that activity based modeling system. Another drawback of these methods are that they cant be
used for location choice models. Since one primary objective of including these models in the
system is to provide sensitivity to transportation investments and policies, the use of the nonparametric modeling methods is not recommended.
608
Page 37
37
Most of the long-term choice models in use today are based on discrete choice methods with all
the advantages and disadvantages that come along with them. These models are flexible in terms
of modeling structures and variables and can represent non-measurable variables through the
alternative-specific constant. They also provide a sound economic backbone to the analysis and
are sensitive to transportation investments and other policies. The drawbacks of discrete choice
models are the data necessary for estimation and calibration, which may be a limiting factor for
some of the modeling components.
609
Page 38
Yes
Employment industry
Usual workplace
Usual workplace
Home
Choice I
Choice 1
Home
Not home
Not home
N1
NI
38
Discrete choice models can be seen as a series of hierarchical choices--either as sequential model
components where as nested, or as simultaneous model components. In this example, the initial
choice is whether to participate in the labor force, or to get a job and then to choose an industry.
Once an industry is chosen, the next choice is whether to work at home or outside the home and
finally to choose a workplace location. If this model were a nested model, it would have four
levels or it could be developed as four individual multinomial logit models. Current practice for
long-term choice models is to develop a sequential set of multinomial logit models because the
estimation process is more straightforward.
610
Page 39
q k yq* q k 1
Where
q = An index for household
= The latent propensity of households vehicle/car ownership level
= Observed households vehicle/car ownership level
k = An index for the number of vehicle/cars in a household (k = 0, 1, 2, 3)
xq =A vector of exogenous variables
= A corresponding vector of coefficients
= An error term (standard normal or Gumbel distributed)
= The lower bound threshold for vehicle ownership level k
yq =0 (No car)
yq =1 (1 car)
yq =2 (2 cars)
39
611
Page 40
40
Another aspect of long-term choice models is the potential to represent joint choices. This is
theoretically appealing because many choices cannot be sequentially determined and are in
reality jointly determined by people. For example, does one choose their workplace based on
where they live or choose their residence based on where they work? When someone is young,
they may choose their workplace first, and then lived nearby, but as one gets older, they may
choose their home first, and then get a job nearby. In addition, some families may choose their
residence first, and other families may choose their workplace first. Having a joint choice for
residential and workplace location would solve this problem. Nonetheless, these
multidimensional choice models are more complicated to estimate and apply and may therefore
be less practical to include. If an agency sees a need to represent these joint choices in order to
improve the explanatory power of the model, then the additional complexity will be worth it.
612
Page 41
41
One example of a joint choice model was developed in London to predict the joint choice of auto
ownership and transit pass holding. This first chart shows the results of the auto ownership
component for different population segments. Households in urban areas and with low income
are more likely to have no cars and those with employer parking benefits or high income are
more likely to have more cars, as expected. Also, households with kids are less likely to have no
cars. The second chart shows the results of the transit pass component of the model, with seniors
and people with employer transit benefits most likely to have a transit pass and males least likely
to have a transit pass.
613
Page 42
Ordering of decisions:
E.g., Did a household move to the neighborhood because of a
new job, or did the workplace choice follow the choice of
residence or was school location/quality the deciding factor?
42
One of the bigger challenges in multidimensional choices is how to sort out endogeneity issues.
For example, do people choose to own a car because they live in the suburbs or do they live in
the suburbs because they want to own a car. These types of self-selection biases can be difficult
to sort out. In a set of sequential models, these decisions must be ordered and it may be
challenging to determine an order that works for all people. Nonetheless, the practical needs for
activity-based modeling require a sequential approach, so the ordering of decisions should be set
to represent local conditions as much as possible.
614
Page 43
43
Cross-nested choice models is a form of random utility model that allows flexibility of assign
alternatives to nests across alternatives. The cross nested modeling structure is appealing because
it can capture more complex situations where correlations among alternatives can be
acknowledged.
615
Page 44
Gasoline
Hybrid
Electric-Plugin
Hybrid-Electric
44
An example of a cross-nested model is for vehicle purchase and use decision models, typically
used for estimating energy consumption and emissions. These models estimate make, model,
engine type, and fuel efficiency of new vehicles. Typically, the cost and performance of the new
cars are included as well as attitudes of the buyers for certain makes and models, as well as price
ranges and features. The vehicle purchase decision is clearly a long-term choice and the vehicle
use decision may be both a long-term and a short-term choice, depending on whether each
person wants to drive their own car or whether people share a car or whether people swap cars
for different purposes. This example is of a vehicle choice model developed for the California
Energy Commission.
616
Page 45
Compact
Car Nest
Compact
Gasoline
Compact
Hybrid
Compact
SUV Nest
Gasoline
Nest
Compact
Flex Fuel
Hybrid
Nest
SUV
Gasoline
Flex Fuel
Nest
SUV Hybrid
SUV Flex
Fuel
45
Here is the cross-nested vehicle choice modeling structure developed for the California energy
commission to identify initially the type of car and then the fuel efficiency of the car. The cross
nests allow for different types of cars and fuel efficiency types at each level.
617
Page 46
Compare
Nested
(NL) and
CrossNested
(CNL)
Logit
Vehicle
Choice
Models
NL (fuel)
no
Nesting parameter
mixture mixture
Standard Gasoline
0.68
0.81
Flex Fuel/E85
0.76
0.88
Clean Diesel
0.82
1
Compressed Natural Gas
0.9
1
Hybrid-electric
0.56
0.45
Plug-in Hybrid-electric
0.74
0.88
Full Electric
1
1
Subcompact car
Compact car
Mid-size car
Large car
Sport car
Small cross-utility car
Small cross-utility SUV
Mid-size cross-utility SUV
Compact SUV
Mid-size SUV
Large SUV
Compact van
Large van
Compact pick-up truck
Standard pick-up truck
-
NL (vehicle)
no
mixture mixture
0.9
0.63
0.72
0.57
0.71
0.59
1
1
1
1
0.77
0.79
0.61
0.69
0.75
0.65
0.68
0.48
0.77
0.67
0.76
0.81
1
1
0.63
0.88
0.71
0.76
1
1
CNL
no
mixture
0.4
0.08
0.8
0.97
0.46
0.6
1
0.81
0.53
0.48
1
1
0.65
0.37
0.66
0.15
0.6
0.4
1
0.47
0.39
1
mixture
0.59
0.08
1
1
0.33
0.79
1
0.21
0.23
0.20
1
1
0.68
0.27
0.34
0.16
0.20
0.45
1
0.87
0.46
1
46
This table is based on the same cross nested legit vehicle choice models presented above, but
shows a comparison with traditional nested logit models developed for fuel types, as well as
vehicle types. This example demonstrates how the cross-nested model structure can provide a
more complete picture than the nested model structure.
618
Page 47
Discrete-Continuous Choices
Discrete allocation of autos to household members and
continuous allocation of vehicle type usage (VMT)
47
Another aspect of the long-term choice models is whether to include discrete, continuous choices
for different model components. In this context discrete allocation of autos to household
members could be combined with continuous allocation of vehicle type usage. Most often,
however, the long-term choices will be modeled as discrete choices since they by definition are
not changing over time. Discrete, continuous choices may be useful for studies of alternative
vehicle technology, fuel efficiency, and greenhouse gas emissions.
619
Page 48
Costs
48
There are many aspects of integrating the long-term choice models within the activity-based
modeling system to consider. The benefits of integration are that long-term choices are sensitive
to transportation investments and policies such as pricing strategies and demand management
programs. Another benefit is that long-term choices can be represented with different
characteristics than short-term choices. There are costs to including long-term choices as
separate modeling components because they add runtime and complexity to the process. In
addition, the feedback between travel times, costs, and accessibility should be equilibrated with
other downstream model components. There is a trend to add more long-term choice models and
hence more sensitivity indicating that these benefits outweigh the costs.
620
Page 49
PLACE3S, PECAS,
residential,
employment, school
enrollment and other
information
49
The long-term choice models are one element of this full activity-based modeling system as
demonstrated in this example from the Puget Sound Regional Council in Seattle. In this example,
PSRC includes usual work and school location choice models in two places: as long-term
choices and as part of the person day activity pattern models. The long-term choice models also
include auto ownership models, usual mode to work models, arrival and departure time to work
models, transit pass models, and paid parking at workplace models. The number of long-term
choice models has been growing steadily over time to incorporate various aspects of long-term
choices and to provide more explanatory power for the travel demand models.
621
Page 50
Mobility Models
Auto ownership/availability model
Free Parking Eligibility
50
Another example of how long-term mobility choice models can be incorporated within the
activity-based modeling is in Columbus at the mid-Ohio regional planning commission. In this
example, the workplace and school location choice models are embedded within the joint daily
activity and travel simulator. The long-term choices represented are the auto ownership and free
parking eligibility models.
622
Page 51
Long-Term Choices
Usual Work and School Location Models
Auto ownership/availability model
Free Parking Eligibility / Reimbursement
Toll Transponder
51
Another example of how long-term mobility choice models can be incorporated within the
activity-based modeling is in San Diego. In this example, the workplace and school location
choice models are long-term choices, along with transponder ownership and free parking
eligibility models.
623
Page 52
Auto
Transit
Non-Motorized
52
This is a unique although simple analysis that we first implemented in Jerusalem, Israel as part of
the ABM development project. The population in Jerusalem for this analysis was broken into
three groups: Arab, orthodox Jewish, and secular Jewish since they are characterized by very
different travel behavior. In each group, workers were broken into two subgroups; those who
have a reserved or paid parking at work and those who dont. For each subgroup, modal split was
calculated with respect to the non-work trips. There is a strong correlation between reserved
parking at work and car orientation in modal split for non-work trips, although seemingly these
choices are unrelated. This is a strong manifestation of modality captured by mobility attributes.
624
Page 53
53
Commuting to work/school is normally the most frequent trip made by workers and students.
Mobility attributes (car ownership, transit pass) are largely defined by commuting conditions.
Also, modality style is largely formed by commuting. Further on, mode choice for other trips is
largely driven by mobility attributes and modality. Inclusion of this sub-model in the model
system enables this important linkage of mode choice decisions across different trips made by
the same individual. This important effect cannot be incorporated in a conventional 4-step
model.
625
Page 54
Approach to Forecasting
Jerusalem ABM has a special extended sub-model for
mobility attributes:
Modeled with interactions and trade-offs between different
mobility attributes
Sensitive to socio-economic, demographic, and travel
variables
Allow for scenario analysis, policy levers, and dynamic trends
through adjustment of alternative-specific constants
54
The Jerusalem ABM includes an example of a special extended sub-model for mobility attributes
with more details included and somewhat more sophisticated model structure applied. Mobility
attributes are modeled with interactions and trade-offs between different attributes. The system is
designed to be sensitive to socio-economic, demographic, and travel variables It allows for
scenario analysis, policy levers, and dynamic trends through adjustment of alternative-specific
constants.
626
Page 55
Mobility
attributes
Vehicle Holdings
Information
55
The Jerusalem model for mobility attributes includes three sub-models applied iteratively with
inter-linkages between them rather than sequentially. The first model represents a joint model of
many mobility attributes applied for each person in the household separately including a role of a
usual driver of car (i.e. person need in a car). The second model ingrates person needs within the
households with respect to number of cars and car type by body and fuel (7 car types). The third
model allocated cars to usual driver within the household. These three models are iterated several
time to integrate the associated choices.
627
Page 56
Telecommuting
Fuel price and taxation
Toll transponders and other advanced toll collection methods
56
A challenge in developing models to address new policies and technologies is that there are no
observed choice data sets by definition, and so stated preference surveys are needed to evaluate
the tradeoffs for these models. In some cases, field experiments and pilot studies of
implementations or demonstrations of these new policies and technologies can provide insight on
travel behavior for these choices. For example, we have collected before and after surveys in
Seattle and Atlanta on travel with and without tolls in specific corridors. These surveys also
included questions about workplace locations and auto ownership and could be used to evaluate
the sensitivity of these choices to pricing.
628
Page 57
Data Sources
Estimation Datasets
Combined with
network accessibility
measures and landuse data
Validation Datasets
57
Most of the data required by the long-term mobility choice models will come from the activitybased modeling. Additional data for estimation of these choice models is required and would
typically come from a local household survey data set or the national household travel survey
(NHTS), or other focused, stated preference experiments. These surveys would need to be
supplemented with network accessibility measures and log-sums in a similar way to the travel
demand choice models. In addition, questions about usual mode, departure time, arrival time for
work trips and usual work and school locations are needed to estimate these long-term choice
models, and these questions may not have been included in past surveys if these long-term
choice models were not envisioned as part of the process.
The data required for application and validation are very similar to data required by the rest of
the activity-based modeling system and include the American Community Survey PUMS data
set and census transportation planning package. In addition, specialized surveys that are collected
for travel demand management programs may be quite useful in validation.
629
Page 58
58
There are quite a few emerging practices in the long-term mobility choice model area as
demonstrated by the many choices that are being considered or are under development in activity
based models around the country. In addition, there is investigation into multidimensional choice
structures and other econometric methods in research settings. And, as is the case in many areas,
these models are changing to reflect new and emerging policies and technologies. Another area
that has received considerable discussion is modeling household budgets, which really drives
cost decisions for transportation.
630
Page 59
59
We have covered a wide range of possible long term and mobility decisions that could be
included in the activity-based models. They include the most commonly used models for
workplace and school location choices and vehicle availability models, as well as newer models
under consideration at many MPOs for personal and worker mobility decisions.
631
Page 60
60
These mobility decisions are important because they add explanatory power to the travel models
and provide sensitivity for policy variables of interest to MPOs. These include travel demand
management programs, pricing strategies, and the influence of transportation investments on
longer-term choices for work and school.
632
Page 61
61
These long-term and medium-term mobility decisions can be integrated into the activity-based
travel modeling system in several ways; we have discussed advantages and disadvantages of
various approaches to consider. In addition, we have talked about a series of alternative modeling
formulations that have been considered for mobility models that have been used in various other
venues and in research settings. These may represent the next generation for mobility models.
633
Page 62
634
Page 63
February 2
February 23
March 15
April 5
April 26
May 16
June 7
June 28
July 19
August 9
August 30
September 20
63
Once again, here is the schedule for the webinar series. Our next webinar, three weeks from
today, will cover activity pattern generation models.
Thank you!
635
Page 64
64
636
637
for statewide models, the individual attributes are still relevant, but the coefficients would be
different. The same coefficients should not be transferred, but the mechanics are transferrable.
How are transaction costs handled in the model, such as traveler information?
Maren: There are a few examples of incorporating real-time information into mobility decisions,
for example, comfort, reliability, real-time information, station amenities into mode choice
models.
Are any of you familiar with GPS data used in California or other states, and is it being used in
model estimation?
Peter: Yes, SCAG collected GPS survey data, and at the statewide level for Caltrans, and for
SFCTA in San Francisco. It is being used, primarily to identify the under-reporting bias in
household surveys. A smaller household sample includes GPS data, and that data is analyzed to
determine the aspects of trip under reporting (frequency, length, etc) and then the main survey is
adjusted to correct for the under-reporting. Also the following surveys were 100% GPS (or close
to it): Cincinnati, Cleveland is starting soon, and Jerusalem, Israel. When you compare the
quality and quantity of data between GPS and non-GPS data, GPS surveys are very costeffective. The quantity and quality of travel data is much higher in GPS surveys.
For the ABMs that were developed for San Joaquin Valley do they use both parameteric and
non-parametric methods? Do ABMs use both parametric and non-parametric methods?
Maren: San Jouquin Valleys population synthesizer is largely non-parametric, while all of the
long-term, mobility, and daily travel models are parametric. In any model, there are often
combinations for both parametric and non-parametric methods. In a non-parametric approach,
we try to apply probabilities to discrete groups of people (market segments), which reflect
observed distributions but do not vary according to covariates. For example, an auto ownership
model that is driven entirely by constants is just a way of using logit math to calculate a
probability distribution that can be easily specified by a non-parametric distribution. It is
generally always advantageous to use parametric models so that we can reflect changes in
outcomes due to changes in continuous input variables. But it is not always possible to do so
for example, there may be cases where we dont observe enough cases where we can estimate a
parametric model.
638
639
Page 1
Activity-Based Modeling
Session 8: Activity Pattern Generation
640
Page 2
Acknowledgments
This presentation was prepared through the collaborative efforts of Resource
Systems Group, Inc. and Parsons Brinckerhoff.
Presenters
Peter Vovsha and John Gliebe
Moderator
Stephen Lawe
Media Production
Bhargava Sana
Resource Systems Group and Parsons Brinckerhoff have developed these webinars
collaboratively, and we will be presenting each webinar together. Here is a list of the persons
involved in producing todays session.
Peter Vovsha and John Gliebe are co-presenters. They were also primarily responsible for
preparing the material presented in this session.
Stephen Lawe is the session moderator.
Content development was also provided by Joel Freedman. John Bowman, Mark Bradley and
Maren Outwater provided review.
Bhargava Sana was responsible for media production, including setting up and managing the
webinar presentation.
641
Page 3
February 2
February 23
March 15
April 5
April 26
May 16
June 7
June 28
July 19
August 9
August 30
September 20
3
This is the 9th session overall and the 5th technical session. Understanding of the previous topics
is essential since we will be using the terminology introduced in the previous sessions. In
particular, we assume a basic knowledge of micro-simulation principles as well as familiarity
with tour structures.
642
Page 4
At the previous webinars we discussed overall structure of activity based modeling and
corresponding advantages over 4-step models. We discussed population synthesis and now the
daily activity pattern (DAP) model can be applied to each individual person and households in
the synthetic population. The DAP model uses accessibilities, long-term and mid-term choices as
important explanatory variables. These variables are assumed known when we discuss DAP.
Today we start discussion on how individual travel choices are made on given (modeled). Our
material is limited to a regular weekday. Weekends can be modeled with a similar approach but
many details would be different. The Individual DAP is a central concept of activity based
modeling. It roughly replaces trip generation step pertinent to 4-step models. The DAP model
generates activities, tours, and trips with cross-impacts on each other. DAP is the cornerstone of
activity based modeling and key differentiating feature from 4-step. It is also a good litmus test
on understanding the entire activity based modeling concept.
643
Page 5
Learning Outcomes
Role and placement of DAP model in activity based modeling
Structure of DAP choice model and alternatives in the choice set
Advantages of DAP vs. traditional trip and tour generation
models
How integrity of DAP can be achieved:
For each person, between number of activities, trips, and tours for
different purposes
Across household members, including joint activity and travel
644
Page 6
Outline
Basic terminology
Definition of DAP
Role and placement of DAP in activity based modeling,
linkage with the other models
Relation of DAP to trip and tour generation models in 4step framework
Individual DAP (IDAP) implemented for each person
independently
Coordinated DAP (CDAP) implemented for all household
members
Ongoing research, main directions, and challenges
Activity-Based Modeling: Activity Pattern Generation
Basic terminology
Definition of DAP
Role and placement of DAP in activity based modeling, linkage with the other models
Relation of DAP to trip and tour generation models in 4-step framework
Individual DAP (IDAP) implemented for each person independently
Coordinated DAP (CDAP) implemented for all household members
Ongoing research, main directions, and challenges.
The discussion of these practical examples is the most essential part of this session. The rest is
the necessary build-up.
645
Page 7
Terminology
Main units of activity based modeling analysis:
Activity episode
Trip
Tour
Sub-tour
Primary activity on tour
Half-tour by direction (outbound, inbound)
Daily activity pattern
We will be using activity based modeling terminology intensively. We assume you are familiar
with the main units of activity based modeling analysis from the previous webinars. Today we
will be frequently using classification of activities that has also been discussed before. We will
also introduce and discuss in detail definitions of terms specific to daily activity patterns, with an
emphasis on operational models of individual and coordinated daily activity pattern types.
Activity episode is an event in which an individual is engaged in an activity at a specific place
and time. In modeling out-of-home activities, we usually assume that the activity purpose does
not change while the person remains at the same location and only model changes in activity
when there is a change in location. We typically group all in-home activities together, except
perhaps for modeling in-home paid work.
Trip refers to travel between the locations of activity episodes, including between home and outof-home activity locations.
646
Tour is a sequence of two or more trips, traveling from an anchor location (usually home) to one
or more activity locations (out of home) and with a return trip back to the anchor location,
thereby completing the tour.
647
Page 8
Classification of Activities
Type/purpose
mandatory, maintenance, discretionary
Location
at-home vs. out-of-home
Intra-household interaction
individual, joint, allocated
It is useful at this point to review how we classify activities. Much of this has been covered to
some extent in previous webinars in the series, particularly Webinar 4, which was the
introduction to activity-based modeling frameworks. One way to classify activities is by type and
purpose. We will talk about grouping activity purposes by whether they are mandatory,
maintenance, or discretionarylabels intended to convey prioritization.
We will also classify activities by whether they take place at home or out-of-home. Typically, we
model only out-of-home activities for the purposes of travel, but it is now becoming an accepted
best practice to at the very least model in-home work activities for the purposes of modeling
telecommuting policies. As I just mentioned above, we will also be establishing primary activity
types, destinations and modes on a tour, and of course identifying secondary activity stops.
Finally, we will be discussing in great detail concepts related to intra-household interactions.
This includes joint activities and travel between individuals, often eating out, social/recreational
and shopping; providing rides to other household memberspick-ups and drop offs, especially
children; and allocation of activities. Allocated activities may include activities such as shopping,
648
649
Page 9
We can also talk about daily activity patterns in different ways. Identification of daily activity
patterns begins by classifying activities, tours, trips and stops by purpose and tabulating their
joint frequencies. For example, in a sample population, what is the joint distribution, for a given
day, of the numbers of home-based tours, as classified by mandatory, maintenance and
discretionary purposes? We will show you examples, but typically the simplest daily patterns are
the most commona single tour for the day, with more complex combinations of multiple tours
and sub-tours being much less common. Also, know that persons who stay home all day are
fairly common, particularly among pre-school-age children and adults past the retirement age.
In this webinar we will focus primarily on two ways of defining daily activity patterns
individual and coordinated. An individual daily activity pattern describes a single persons day of
activity and travel, usually defined by tours of particular purposes. A coordinated daily activity
pattern refers to a single household pattern, which implies the joint occurrence of individual
patterns for each household member. The coordinated pattern may also include interactions that
are not included in an individual daily pattern approach, such decision models related to
650
651
Page 10
Maintenance:
Escort Passenger(s)
Shopping
Personal Business (e.g., Medical)
Discretionary:
Eating out
Visiting relatives and friends
Social/Recreational
Activity-Based Modeling: Activity Pattern Generation
10
Classification of out-of-home activities and travel purposes by major types is essential. We have
tended to group these into mandatory and non-mandatory activities. Mandatory activities include
work, business-related and school activities. Maintenance activities include escorting passengers
(primarily children), shopping, medical, etc. Discretionary activities include eating out, visiting
relatives and friends, and social/recreational activities. There are some grey areas between
maintenance and discretionary activities (for example, visiting a shopping mall can combine
shopping and eating out). In addition, the more we learn about these activities the less clear it is
whether people really prioritize maintenance activities over discretionary activities. For example,
a recreational activity might be the big event of the dayand people may make a quick stop at a
store to pick up some food or drinks on the way.
652
Page 11
Population
Long-Term
Mobility
Modes
Space
Activities
Tour Patterns
Time
11
Lets consider how decisions regarding definition and modeling of Daily Activity Patterns
(DAP) affect other aspects of model system design.
This diagram serves as a backdrop for describing the relationships between key design elements
in activity-based modeling. These elements include: defining the population, modeling longterm and mobility-related choices, defining activity types, defining modes, defining tour patterns
and an entire day-pattern elements, as well as the treatment of space and accessibility and
treatment of time. We discussed each of these design elements in the previous webinar on
activity-based modeling frameworks and techniques, and have already devoted entire sessions to
population synthesis, the treatment of space, and to long-term choices and mobility models.
DAPs are defined by activity purposes and by tour patterns. As you shall see in the rest of this
session, DAPs represent how individual persons and groups of persons in households arrange
their days into tours for various activity purposes. We model these patterns directly in part to
recognize that tours made by the same person on the same day are in fact interdependent and that
certain combinations of tours and activities are more likely to be observed than other
653
654
Page 12
Build Alternative(s)
12
Lets consider a transportation planning and policy project that might be faced by an MPO or
DOT and how daily activity pattern modeling fits into the picture. We have used this example in
several previous sessions to talk about how activity-based modeling components might affect the
analysis. Here, we have a planning or policy case in which an agency is looking a bridge
crossing study. Some of the design options that they will need to consider in their alternatives
analysis include:
The bridge will be expanded from 4 to 6 lanes. The two new lanes are expected to be HOV
lanes, one in each direction, and the bridge will be tolled.
Various tolling schemes will be considered, including flat rates and time-variable pricing.
The times of enforcement for the HOV lanes are likely to be peak periods, but another option
is that they be enforced at other hours.
In addition, the regional transit system is simultaneously considering a fare policy that will
charge more during peak periods.
655
Page 13
13
For this bridge example, daily activity pattern generation may be affected in the following ways.
Persons whose accessibility is improved by the bridge may increase their discretionary activity
episodes. In order to take advantage of HOV lanes, household members may decide to share
rides to work. This not only affects mode choice, but also affects daily activity patterns, because
of the coordination of patterns between household members. It may even lead to increased nonwork joint activities.
656
Page 14
Mobility
Choices
Model
Inputs
Long-Term
Choices
Daily Activity
Pattern
Tour
Details
Transport
level-of-service
and
accessibilities
Trip
Details
Network
assignments
Model Outputs
14
Activity based models vary in terms of their exact structure. However, there are some common
elements. All models start with some inputs (transport networks, land-use forecasts, etc). A
synthetic population is created. Then long-term choices (work location, school location) are
typically modeled. Mobility choices (such as auto ownership) are modeled next. Daily activity
patterns and tours are created for each person in the synthetic population, and details about each
trip are filled in. Trips are assigned to networks, and transport levels-of-service are fed up
through the system.
DAP is a first major travel-related model in the activity based modeling chain that generates
activities, tours, and trips (analogous to trip generation in 4-Step). There are some variations in
advanced activity based models w.r.t. the DAP model structure that we will be discussing today.
They primarily relate to the incorporation of intra-household interactions.
657
Page 15
15
Despite the common view on activity based models in general and DAP in particular as a
revolutionary concept it was actually largely a long evolution process that went through four
major steps. First two steps relate to conventional 4-step models. The 3rd step relates to advanced
4-step models and simplified activity based models. The 4th step corresponds to advanced
activity based models that we discuss today. The 3rd and 4th types of models specifically tailored
to by applied in a micro-simulation activity based model. Conceptually DAP is just a
simultaneous or joint trip/tor/activity frequency choice model for all purposes.
658
Page 16
Home
Work
Shop
Home
Work
Shop
16
The first important technical point to discuss today is why we want to model trips for different
purposes jointly in the DAP framework? Whats wrong in applying several separate models for
different trip rates? After all it is simpler. The reason for this is that trip rates for different
purposes are not independent for the same individual. For example, a person who did not have
either mandatory or maintenance trips on a day would have a higher probability of not having
discretionary trips either (being sick, absent from the town, working from home, etc). However,
the same logic is reversed when it comes to comparison between one and multiple activities. For
example, those who have multiple maintenance activities on the day will have a lower
probability of multiple discretionary activities on the same day. Thus modeling different trip
purposes for the same person independently loses information and may result in an illogical,
inconsistent DAP.
659
Page 17
Children
Mandatory
School
School
University
School
University
Work
At Work
Discretionary Maintenance
Workers
M
M
M
D
D
17
This represents an historical attempt of transitioning from trip generation (3rd type) to DAP
models. It was applied in the NY activity based modeling designed in 2001. In this model DAP
is broken into a subset of 13 interlinked tour-frequency models for different purposes and person
types with some simplified accounting for intra-household interactions.
660
Page 18
Daily
Activity-Travel
Pattern
At
Home
Pattern type
Primary
Activity
On
Tour
Primary
Destination
Primary Tour
Configuration
Secondary
Stops
At
Home
Secondary
Activities
On
Tour
Primary
Destination
Secondary
Stops
Secondary Tours
Configuration &
Sequencing
18
Definition of J. Bowman and M. Ben-Akiva per Bowmans masters thesis, where the first
attempt was made to formulate an integrated individual DAP for a person instead of a set of
frequency models. This formulation distinguished between primary and secondary activities and
subsequently between travel tours. We will discuss some more recent modifications of this
model in detail today.
661
Page 19
19
We start with a concept of Individual DAP applied for each person. At this point we do not
consider intra-household interactions explicitly and focus on one person. There are three major
components of IDAP that we want to model:
Pattern type (main characteristic of entire day; most important determinants of person travel
behavior; strongly constrains generation of tours by purpose);
Set of tours (by primary activity and primary destination); and
Secondary activities (stops) on the way to and from primary destinations.
662
Page 20
Age
Full-time worker
18+
Full
Part
Part-time worker
18+
Part
Part
University student
18+
Part
Full U
Non-worker
18-64
Retired
65+
16+
6-15
Full S
Pre-school child
U6
Full S
Part
Part
Full S
20
Typical DAPs are very different for different person types. When we analyze DAP stats we will
consider 8 main person types as defined in the table (read the definitions). Black attributes are
necessary. Grey attributes are optional.
663
Page 21
Mandatory
At least one out-ofhome mandatory
activity (tour) and
any other activities
Non-mandatory
No mandatory activities; at
least one out-of-home nonmandatory activity (tour)
21
Every DAP falls into one and only one category. This is obviously is the most important daylevel decision made by each person that has a crucial conditional impact on the number and
schedule of all activity episodes. We start our analysis at this aggregate level by classifying
DAPs by whether they are at-home (no travel), mandatory, or non-mandatory. As defined above,
a mandatory pattern would include at least one out-of-home mandatory activity (tour)work,
college, and schooland may include any other non-mandatory activities. The non-mandatory
DAP group excludes mandatory activities and must include at least one non-mandatory activity
(tour)shopping, eating out, social/recreational, personal business/medical, and providing rides
to other household members (escort).
664
Page 22
Retired Sch-drive
Non-mandatory
Schpredr
Preschool
Home
22
Observed frequency of DAP types is obviously a strong function of the person type. Full-time
workers, university students, and school children are naturally characterized by a dominance of
the Mandatory DAP type. Contrary to that, non-workers and retired people are characterized by a
dominance of the Non-mandatory DAP type as well as frequent staying at home.
665
Page 23
Retired Sch-drive
Mandatory
Non-Mandatory
Schpredr
Preschool
At Home
23
This distribution is stable across regions and years since it reflects on the most fundamental
features of travel behavior. However, there are some significant differences and sensitivities to
the local conditions. For example, the work attendance factor may vary between 70% (Bay Area)
and 80% (San-Diego) that has significant implications on regional VMT, etc.
666
Page 24
0 tours
1 work
2+ work
1+ work & 1+ school/university
1+ school/university
0 tours
1+ shopping/escort/maintenance
1+ eating/visiting/discretionary
1+ shopping/escort/maintenance & 1+ eating/visiting/discretionary
IDAP
Home
1
Activity-Based Modeling: Activity Pattern Generation
Mandatory
44=16
Non-mandatory
3
24
Now lets consider a realistic but simplified example of an IDAP structure where we limit the
possible frequency of mandatory and non-mandatory tours as shown on the slide. We also will
focus on number of tours and assume that there are no intermediate stops on these tours. Thus all
tours have a form of a round trip with a single destination. We will also distinguish between 2
aggregate non-mandatory purposes only (maintenance and discretionary) without details of each
particular purpose. With these simplifications we arrive at 20 possible IDAPs of which one
pattern is to stay at home all day; 16 patterns include at least one work or school activity; and 3
patterns have only non-mandatory tours.
667
Page 25
Daily
Activity
Patterns
DAP Type
Mandatory Tours
Non-Mandatory Tours
Home
Mandatory
1 Work
1+ Escort/Shop/Maintenance
1+ Eating/Visit/Discretion
1+ Escort/Shop/Maintenance
1+ Eating/Visit/Discretion
2+ Work
1+ School/University
Non-Mandatory
DAP alternative
10
1+ Escort/Shop/Maintenance
11
1+ Eating/Visit/Discretion
12
13
14
1+ Escort/Shop/Maintenance
15
1+ Eating/Visit/Discretion
16
17
1+ Escort/Shop/Maintenance
18
1+ Eating/Visit/Discretion
19
20
25
Exploring the combinatorics of DAP are essential for understanding this webinar. Thus, we list
all these patterns explicitly for this example. You can see that the home pattern does not have
any travel. All mandatory patterns are different but each of them has at least one mandatory
activity. Please note that Mandatory pattern can also have non-mandatory tours although not
necessarily. All non-mandatory patterns are different but each of them has at least one nonmandatory activity and does not have any mandatory activity. The list of patterns is mutually
exclusive and collectively exhaustive under the adopted simplification scheme. It includes all
possible combinations.
668
Page 26
Mandatory Tours
Home
Mandatory
1 Work
2+ Work
1+ Work & 1+School/University
1+ School/University
NonMandatory
7.4%
60.8%
3.1%
1.2%
3.1%
24.5%
26
All listed patterns are observed in reality for workers chosen as an example of person type
(BATS, 2000 in this example). You can see that while the distribution is logically dominated by
a Mandatory pattern with a conventional single work tour there are enough nonconventional
cases. The relatively high share of workers having a Non-mandatory pattern on a regular
weekday is a signature feature of the Bay Area with a large share of telecommuters and workers
with flexible workdays.
669
Page 27
58.7%
22.2%
14.0%
5.2%
27
If we focus for a moment at the relative frequency of non-mandatory tours only we obtain the
following logical distribution for workers. Majority of cases logically corresponds to a single
non-mandatory tour. Overall maintenance activities are more frequent for workers on the regular
workday compared to discretionary activities.
670
Page 28
0%
Activity-Based Modeling: Activity Pattern Generation
Independent
Observed
50%
28
Now we present the relative observed frequency of each of the 20 IDAPs. We also present a
frequency calculated as a Cartesian product of the observed frequency of mandatory activities by
observed frequency of non-mandatory activities. This calculation assumes independent tour rates
for mandatory and non-mandatory tours. You can see that the independent calculation is quite far
from the observed frequency. This means that IDAP cannot be predicted as a combination of
frequencies by purpose. In other words, tour rates for different purposes are not nearly
independent. This is another important illustration why the concept of integrated IDAP is
essential for understanding travel behavior.
671
Page 29
W1 (pattern) +
W (tour) + M (tour)
DAP=13
(Work + School +
Maintenance + Discret)
WS (pattern) +
W (tour) + S (tour)
M (tour) + D (tour)
N (pattern) +
D (tour)
DAP=19
(Discretionary)
29
To model choice of IDAP we have to form a meaningful utility function that portrays the worth
of each pattern. This utility is formed in a component-wise fashion reflecting that each activity
has a certain value for each person plus there is an additional effect of the intensity of the entire
pattern. Several examples of utilities are shown on the slide. This technique where a utility for
each IDAP is combined of these components is absolutely essential when a real IDAP choice
model is formed with thousands of possible patterns. This is a way to handle a large
combinatorial choice with a parsimonious structure in terms of coefficients. This technique is
also frequently applied for other sub-models of an activity based modeling like TOD choice that
will be discussed at the next webinar.
672
Page 30
30
Now we are ready to consider an example of real IDAP applied in an advanced operational
activity based modeling in practice with all details. We have chosen the Sacramento activity
based modeling that is called DaySim as an example. Similar structures were successfully
applied in several other activity based models in practice. Overall, it is a well-tested and wellexplored structure.
673
Page 31
31
The Sacramento IDAP model predicts for each person, tours by purpose, and purposes for which
intermediate stops occur during the day. IDAP also offers a high level of intra-person
consistency of the days tours and stops for 7 different purposes. It is relatively simple and
possible to enumerate all main IDAPs, although results in thousands of alternatives. Intrahousehold interactions not modeled explicitly but somewhat accounted implicitly through
household variables.
674
Page 32
Model
applied after
TOD, DC,
mode
Main pattern
configuration
Exact daily number of
tours
32
The DaySim IDAP goes through the following foru steps. First three steps are applied
sequentially before the time-of-day, destination and mode choices are modeled for each tour. The
last, 4th step is modeled after the times-of-day, destinations and modes have been predicted for
each tour. Thus, there are some additional sub-models between the 3rd and 4th step that we skip
over today. Those will be covered in subsequent webinars on those topics.
675
Page 33
Possible # tours
1=Work
1+
1+
2=School
1+
1+
3=Escort
1+
1+
4=Personal business
1+
1+
5=Shopping
1+
1+
6=Eating out
1+
1+
1+
1+
33
The first component of the IDAP relates to the main pattern configuration in terms of tours and
additional stops on these tours. At this step we do not yet model details like an exact number of
tours or stops if more than 1. However, for each of the 7 purposes we want to know if there is at
least one tour made and at least one additional stop made. Even with these simplifications, the
combinatorics of the IDAP model are impressive. If we consider all possible combinations
mechanically we arrive at over 16,000 possible different patterns; however, the number of
possible combinations can be effectively truncated since many of them are never observed (for
example a pattern with all seven tour purposes present for the same person on the same day is
unrealistic). This reasonable truncation leads to about 2,000 of choice alternatives that can be
handled efficiently as a simultaneous choice.
676
Page 34
34
Even after truncation we have way too many alternatives to construct a different utility for
eachwhat to do? The solution that is a very important technique applied in activity based
models that we already have touched upon is a parsimonious component-wise structure of the
utilities. We have many utilities but they are combined of a limited number of components and
require only a limited number of coefficients to estimate. The main utility components of IDAP
are:
Nominal utility component for tour (Tx) and stop (Sx) by purpose (x);
Tour & stop frequency related components (NTx, NSx, NTx+NSx) by purpose; and
Interaction terms ensuring intra-person consistency and trade-offs between tours and stops by
purpose (ITxy, ISxy, ITSxy) by pairs of purposes.
677
Page 35
Tour component
Work tour +
shopping tour
(work) + (shop)
Work tour
w/shopping stop
(work)
(shop)
(work, shop)
(work) + (school)
(recreational)
(work, school) +
(work, recreational) +
(school, recreational)
Recreational tour
(recreational)
2 shopping tours
(shop) + (shop)
Stop
Component
Interaction term
(work, shop)
35
In this table we provide several examples how utility can be constructed for different patterns
and what are the main components. You can see that many components are used in several
IDAPs. For example, the work tour utility component is included in first three patterns.
678
Page 36
Form a utility by components for IDAP that includes work tour and shopping tour:
Component that measures the utility for total number of 1+ tours
Component that measures utility for 1+ tours for a particular purpose
Component that measures the joint utility of 2 tours of different purpose
Where:
36
Now lets consider a more specific example. How we can form a utility function for a pattern
that includes one or more work tours and one or more shopping tours? We need to include a
component for each tour type and also account for interactions. Thus our utility would include
the following 4 components as shown in the formula. Each of these components is in itself a
function that can include a constant, some household and person variables, as well as some
accessibility variables.
679
Page 37
1 tour
2 tours
3 tours
Constant
-7.469
-14.18
Accessibility to shops
0.5011
0.9517
Full-time worker
0.5642
0.5642
-0.6396
-0.6396
-1.661
-1.661
1.105
1.105
0.5436
0.5436
Only adult in HH
0.5404
0.5404
HH Income 75K+
0.3538
0.3538
Work at home
0.4937
0.4937
# work tours
-2.443
-2.443
-0.2152
-0.2152
37
After the main pattern configurations have been predicted for each person, we have to add details
for each tour. It might be a single tour that is the most frequent case, but some people implement
2 or even 3 tours per day of the same purpose. A reminder is that tours might be non-motorized
and short. The exact number of tours is defined by a choice model that considers numbers of
tours as alternatives as shown in the example for shopping tours. A single tour per purpose is the
base alternative with 0 utility. Alternatives that represent multiple tours have many explanatory
variables. Large negative constants reflect on the fact that having more than 1 tour for the same
purpose is an infrequent case.
In this model formulation, many coefficients (in black) proved to be generic across alternatives
since they relate to the persons propensity to be engaged in multiple shopping activities, while
the constants and impact of accessibilities (in red) proved to be the differentiating part.
680
Page 38
Choice
1+
1st purpose
2+
2nd purpose
3+
3rd purpose,
etc
38
Finally, we have to define number of stops on each tour. We have a separate choice model for
stop purpose that is applied by half-tours. It is a so-called sequential stop-and-go structure. It is
controlled by the total number of secondary stop purposes predicted by the main IDAP
configuration that we discussed before.
681
Page 39
39
682
Page 40
Columbus, OH (MORPC)
Lake Tahoe, NV (TMPO)
Atlanta, GA (ARC)
Bay Area, CA (MTC)
San Diego, CA (SANDAG)
Phoenix, AZ (MAG)
Chicago, IL (CMAP)
Miami, FL (SERPM)
40
As we discussed before one of the fundamental limitations of the IDAP model is that it predicts
activities, tours, and trips for each person independently of the activities, tours, and trips made by
the other household members. To overcome this limitation a concept of CDAP has been
developed. It is a further development and generalization of the IDAP concept that accounts for
intra-household interactions explicitly. It has been successfully applied in many activity based
models in practice, specifically in activity based models of the CT-RAMP family:
Columbus, OH (MORPC)
Lake Tahoe, NV (TMPO)
Atlanta, GA (ARC)
Bay Area, CA (MTC)
San Diego, CA (SANDAG)
Phoenix, AZ (MAG)
Chicago, IL (CMAP)
Miami, FL (SERPM)
683
Page 41
41
684
Page 42
Daily pattern
Work, School
Non-mandatory
Stay at home /
vacation
Fractional probability:
1st Worker
0.70
0.15
0.15
2nd
0.60
0.25
0.15
0.65
0.05
0.30
Worker
Child
Crisp choices:
1st Worker
Go to work
2nd Worker
Child
Major shopping
Sick at home
42
685
Page 43
43
Lets consider another simple numeric example with the observed numbers that are very stable
and practically identical across all regions. For each full-time worker there is an 80% probability
of going to work and a 20% probability of not going to work. An IDAP model will replicate this
successfully. However, if we consider a households with 2 workers and apply IDAP to each of
them independently we obtain the joint result in which there is a 64% chance of both persons
going to work, a 32% chance of one person going to work, and only a 4% chance of neither
person going to work. This is not what we observe in reality. In reality, in a 2-worker household
in 72% of cases both workers go to work, in 10% of cases neither of them go to work, and in
18% of cases one of the workers go to work, i.e. we observe a coordination between DAPs of
different household members. CDAP will replicate this joint effect while IDAP would fail.
686
Page 44
Episode level:
Shared activity and joint travel (sporting event)
Escorting (children to school)
Allocation of maintenance tasks (shopping,
banking)
Car allocation
44
In general there are many layers of intra-household interactions that are important to consider
when modeling DAP. Some of them manifest itself at the entire-day level:
Staying at home / absent together (vacation, indoor family event, care of sick child)
Non-mandatory DAP together (day-off for major shopping, outdoor family event)
687
Page 45
Home
No out-of-home
activities (tours) or
absence from
home/town
Mandatory
At least one out-ofhome mandatory
activity (tour) and
any other activities
Non-mandatory
No mandatory activities; at
least one out-of-home nonmandatory activity (tour)
45
This is a reminder that every person has a DAP type that falls into one and only one category.
This is obviously is the most important day-level decision made by each person that has a crucial
conditional impact on the number and schedule of all activity episodes. We will be using this
classification but consider several household members together rather than one at a time. We will
see that strong intra-household interactions already show up at this aggregate level.
688
Page 46
School
predriving
School
driving
Univ
student
Full-time
worker
Part-time
worker
Nonworker
46
On this slide, we present relevant stats on sharing the same Home or Non-Mandatory pattern by
several household members. You can see that school children, especially younger ones, are
characterized by a very high degree of sharing. For example, if a preschool child stay at home or
have a non-mandatory travel day (in both cases not going to day care or school) there is a very
high probability for somebody else to share this pattern, i.e. not go to work or school. For adult
household members the probability to share a home or non-mandatory pattern is lower compared
to children but it is still very substantial (40%+).
689
Page 47
50%
40%
30%
20%
10%
0%
Ind
Get
off
Pick Get-in
up
Mid Ohio
47
Another related stats show a share of joint travel by household members. It is very high in
different metropolitan regions. Individual travel constitutes only 50-60%. Almost half of travel
tours are somewhat joint between the household members. Fully-joint tours are the biggest chunk
(~20%),, but all partially joint tours together total to the same 20%. There is of course a
significant variation across different travel purposes with respect to the share of joint travel and
also across household and person types. Thus, it is important to model intra-household
interactions. How can we do that?
690
Page 48
1 person: 3 alternatives
2 persons: 33=9 alternatives
3 persons: 333=27 alternatives
4 persons: 3333=81 alternatives
5 persons: 33333=243 alternatives
48
The main approach is to model DAP choice for all household members together, i.e.
simultaneously rather than sequentially. Each person has 3 alternatives but when we consider
several household members, the number of alternatives becomes quite large but still manageable.
The key factor that we will discuss later is a parsimonious component-wise specification of
utilities. It is important to understand why we have 9 alternatives for 2 persons, 27 alternatives
for 3 persons and so forth. In general, this type of combinatorics is essential for understanding
activity based models.
691
Page 49
Person 1:
Mandatory
Home
Non-mandatory
Person 2:
Person 3:
49
In the CDAP structure each possible choice of each person is combined with each possible
choice for the other household members. Consider an example of a 3-person household. Person 1
has 3 choices. Each of them can be combined with 3 possible choice alternatives for Person 2.
This yields 9 possible combinations between Persons 1 and 2. Now each of these 9 combinations
is combined with 3 possible choices for Peron 3 that results in 9*3=27 alternatives. This
structure is not simple and results in a large number of alternatives for bigger households. Is it
really essential to go into these complexities?
692
Page 50
50
In other words, is coordination of DAPs that significant? Here is another statistical analysis that
is important to understand and recognize.
If intra-household interactions are not significant and DAPs independent across persons,
frequency of any DAP type combination would be close to product of individual frequencies.
Significant biases in group-wise choice frequency versus products of individual frequencies
express intra-household interactions
All possible 36 pair-wise combinations and 120 three-way combinations of 8 person types were
explored with respect to joint NON-MANDATORY and HOME patterns (Atlanta HTS, 2001,
8060 HHs, 2 days).
Here are some of the results.
693
Page 51
15%
10%
5%
0%
Non-work &
Presch
Retired &
Retired
Work-part &
Presch
Presch &
Presch
Work-full &
Work-full
Blue bars correspond to a calculation of a product if individual frequencies. Thats what you
obtain if you apply IDAP. Red bars correspond to actual, observed joint patterns. For example,
logically preschool children and caretaking non-workers stay at home the whole day together.
This is true for almost every other pair of household members.
694
Page 52
52
The differences are even more prominent when we consider triples of household members.
Independent DAP calculation fails to recognize that household members coordinate their DAPs.
695
Page 53
53
Now consider a fully realistic by slightly simplified version of CDAP where we choose 3
representative persons to model, the 1st household head, second household head (if present), and
youngest child (if there is a child in the household). The other household members left aside are
frequently the most individual including older school children, colleague students, or granny
living in. This trio that includes the household heads and youngest child is good example to
illustrate how CDAP works. The other household members can then be added sequentially oneby-one to the modeled three.
696
Page 54
Total:
3+9+9+27=48 alternatives
54
If we limit number of modeled persons in the household to 3 we obtain the following choice
structure for our training purpose. In a 1-person household we model the household head who
has three choice alternatives. In a 2-person household we may have two adults, or an adult plus a
child. In both cases we have to model nine choice alternatives. In a 3-person household we have
2 adult household heads plus child (by virtue of the rule how we choose up to 3 representative
members in each household). A 3-person household requires 27 choice alternatives to consider.
The total is 48 alternatives that we have to consider and develop a utility function.
697
Page 55
55
The utility function for each CDAP alternative has 3 types of components. First type reflects
individual choice preferences (Home, Mandatory, Non-mandatory). We have to form a utility
function for each adult household member that should address such person characteristics as
gender, age, income, worker status, etc. We also have to form a utility for the child that should
address such person characteristics as age, school grade, etc. The second type of utility
component includes 2-way pair-wise interaction terms between each pair of household members
if they choose the same DAP (added utility of joint participation in activities). The third type of
utility term includes 3-way interactions when all three persons choose the same DAP.
698
Page 56
2nd head
Child
Missing
Missing
H1
Missing
Missing
M1
Missing
Missing
N1
Alternative
Utility
56
For a 1-person household, we have only one household head. The 2nd household head and child
are missing. Thus, for each of the 3 alternatives we have a single utility component. H1
corresponds to choice of the Home DAP by the 1st person. M1 corresponds to choice of the
Mandatory DAP by the 1st person. N1 corresponds to choice of the Non-mandatory DAP by the
1st person. Each component, H1, M1, and N2 can include many explanatory variables or utility
terms.
699
Page 57
2nd head
Child
Alternative
Utility
Missing
H1+H2+HH12
Missing
H1+M2
Missing
H1+N2
Missing
M1+H2
Missing
M1+M2
Missing
M1+N2
Missing
N1+H2
Missing
N1+M2
Missing
N1+N2+NN12
57
For a 2-person household with 2 adults, a child is missing. We have 9 choice alternatives for
which utilities are constructed in the following way. First, 2 individual components are including
reflecting the DAP chosen by each person. Secondly we add interaction terms to the 1st and 9th
alternatives where both person choose the same DAP. For the 1st alternative, they both stay at
home. For the 9th alternative, they both have a non-mandatory travel day and do not go to work.
These additional terms make the difference between IDAP and CDAP and express and added
utility of joint participation in implied activities. There is no specific term of joint mandatory
DAP (in the 5th alternative) since mandatory activities like work or school are primarily
individual.
700
Page 58
2nd head
Child
Alternative
Utility
Missing
H1+H3+HH13
Missing
H1+M3
Missing
H1+N3
Missing
M1+H3
Missing
M1+M3
Missing
M1+N3
Missing
N1+H3
Missing
N1+M3
Missing
N1+N3+NN13
58
Following the same logic we can construct utility functions for a 2-person household that
includes an adult and child. However, the variables entering each utility component will be
different from the previous case.
701
Page 59
1st head
2nd head
H
Child
Alternative
Utility
H1+H2+H3+HHH
H1+H2+M3+HH12
H1+H2+N3+HH12
H1+M2+H3+HH13
H1+M2+M3
H1+M2+N3
H1+N2+H3+HH13
H1+N2+M3
H1+N2+N3+N23
10
M+H2+H3+HH23
11
M+H2+M3
12
M1+H2+N3
13
M1+M2+H3
14
M1+M2+M3
15
M1+M2+N2
16
M1+N2+H3
17
M1+N2+M3
18
M1+N2+N3+NN23
19
N1+H2+H3
20
N1+H2+M3
21
N1+H2+N3+N13
22
N1+M2+H3
23
N1+M2+M3
24
N1+M2+N3+NN13
25
N1+N2+H3
26
N1+N2+M2+NN12
27
N1+N2+N3+NNN
M
Activity-Based Modeling: Activity Pattern Generation
N
59
Finally, we can construct all utilities for a 3-person household for all 27 alternatives. The
principle is the same. For each alternative, we sum 3 individual person components, then add
pair-wise interactions where appropriate, and, lastly add 3-way interactions where appropriate.
For example, alternative 10 assumes that 1st household head goes to work while the 2nd adult and
child stay at home. Since this is a frequent child caretaking case, there is an added interaction
term for the 2nds adult and child to stay at home together on the same day rather than on
different days. This is a parsimonious structure where instead of 27 unique utilities we have to
estimate only 9 individual components, 6 pair-wise components, and 2 try-way components.
These components are reused multiple times in different utilities.
702
Page 60
Pre-School Child
Pre-Driving School
Child
-2
Relative utility
4
60
Each component can include many variables. Here is an example of impacts of person age on
certain individual DAPs for different person types. For example, logically, preschool children of
age 4-5 more frequently go to school, kindergarten or day care compared to younger children.
Part-time workers of younger age (U35) less frequently have non-mandatory and mandatory
patterns, etc.
703
Page 61
Female (NM)
Female (M)
-1
3
61
In a similar way, we can capture many effects associated with gender. For example, female
university students and workers have non-mandatory DAP more frequently compared to males.
In a similar way we can analyze impacts of many other variables, like income, car ownership,
density, accessibility, etc. CDAP models applied in practice include hundreds of explanatory
variables encapsulated in the utility components described above.
704
Page 62
SD
RT
NW
US
PW
Pre-school
Child (PS)
Pre-driving
Age School
Child (SP)
Driving Age
School Child
(SD)
Retiree (RT)
Non Worker
(NW)
University
Student (US)
Part Time
Worker (PW)
Relative
strength of
interaction is
proportional
to the size of
the ball
Full Time
Worker (FW)
FW
62
We also include all possible types of interactions between different person types. Relative
strength of interaction is proportional to the size of the ball (in utility units).
For example, for staying together at home for the entire day, the strongest linkages in relative
terms are between school and preschool children (between them) and with the non-workers and
part-time workers (who are the primary child caretakers). However, some statistically significant
interactions manifest itself almost everywhere (for each pair of person types).
705
Page 63
Mandatory
Non-mandatory
Home
63
CDAP is an essential model in the activity based modeling system. It is responsible for such
crucial choice dimension as going to work or school. This model must be well calibrated and
sensitive to policies that we need to model (or at least take into account). In this regard, the
proposed structure is instrumental in practice since it allows for calibration of individual choices
for each person type based on the observed data or policy scenarios. For example, we observe a
growing share of telecommuting and work from home on a regular basis. The share of regular
telecommuters and workers from home has been more than doubled over the last 10 years (from
5% to 10% of workers). To address this tendency we can adjust the corresponding constant for
workers to meet the target share as shown on the slide. In this regard, CDAP calibration is as
easy as calibration of a conventional trip mode choice model.
706
Page 64
64
We normally calibrate CDAP to replicate the observed shares of DAP type for each person type
exactly. DAP models (whether it is IDAP or CDAP) are of crucial importance and discrepancies
at this stage are not allowed. It must be well-calibrated.
707
Page 65
65
This is the corresponding part of the CT-RAMP model system where CDAP is applied. You saw
it at the previous webinars. After the CDAP has been applied for each person, further details are
predicted conditional upon the chosen DAP type. These details include number of mandatory
non-mandatory tours by purpose and type. In particular, we distinguish between three types of
non-mandatory tours:
Fully Joint: All participants engage in same activity/trip sequence (full participation by
members of same household)
Allocated: Maintenance activities that are conducted individually, on behalf of the household
(escort, shop, other maintenance)
Individual: Discretionary activities implemented individually (includes inter-household
ridesharing for non-mandatory activities)
We will discuss some of the key sub-models applied after CDAP type.
708
Page 66
By individual
Individual
Fully-joint tour
Joint outbound
Joint inbound
Drop-off
Drop-off (outbound)
Get off
Pick-up
Pick-up (inbound)
Get-in
Activity-Based Modeling: Activity Pattern Generation
66
We distinguish several types of joint travel stemming from the tour-based modeling technique. In
particular, we will focus first on fully joint tours where all members of the travel party travel
together and participate in all activities. Further on, we will consider an example of escorting
children to school that is a partially joint tour.
709
Page 67
Pick Getup
in
Mid Ohio
67
This tabulation from two very different regions shows a high share of joint travel (40-50%)
where fully joint tours and partially joint tours of different types have approximately equal shares
(20% each).
710
Page 68
No tours
1 tour
2 tours
Shopping
Eating out
Maintenance
Discretionary
Shop/shop
Shop/eat
Shop/maint
Travel party
composition for each
joint tour
Adults
Shop/discr
Eat/eat
Children
Eat/maint
Mixed
Eat/discr
Maint/maint
Person participation in
each party
Activity-Based Modeling: Activity Pattern Generation
Yes
Maint/discr
No
Discr/discr
68
We model fully joint HH tours by a sequence of three choice models each of a them is a discrete
choice model (either multinomial or nested). The first model is a choice of frequency and
purpose of the joint tours. The second model is a choice of the travel party composition for the
tour. It is a trinary choice model that distinguishes between three travel party compositions.
Then, conditional upon the travel party composition we model person participation in each
appropriate party by means of the binary choice model. In many cases, especially for adult and
children parties person participation is predetermined by the HH composition. But there are cases
especially for big households and mixed parties where participation is really non-trivial.
711
Page 69
70%
60%
50%
No Escort
40%
Pure Escort
Ride Sharing
30%
20%
10%
0%
Driving Age
School Child
Driving Age
School Child
69
Another very frequent phenomenon observed in many metropolitan areas is escorting children to
school. For example, based on the data from Phoenix and Tucson we have approximately 50% of
children escorted to school by a parent. There is also a logical pattern in terms of impact of the
child age, and by direction (outbound is somewhat more frequent than inbound). We also
distinguish between two types of escorting: 1=pure escort and 2=ride-sharing. These two types
are modeled differently.
712
Page 70
Symmetry in need
Symmetry in task
allocation
Availability
Availability
Outbound
70
713
Page 71
71
After all tours have been generated for each individual we have to insert intermediate stops in
these tours. This model is called Stop Frequency model and it is similar to the Exact Number of
Stops sub-model for IDAP but we have to consider zero-stop option for CDAP, that was handled
differently in IDAP. The stop-frequency choice model predicts number of stops and their activity
purpose for each half-tour (outbound, inbound).
We consider up to 4 stops on each half-tour for work tours and up to 2 stops on each half-tour for
non-work tours. This is based on the observed stats in many metropolitan regions in the US. This
results in:
Then we assign an activity purpose for each stop. This can be one of 6 non-mandatory purposes
assigned probabilistically and conditional upon the tour purpose and stop order.
714
Page 72
72
It is interesting to compare IDAP and CDAP structures in this particular respect. IDAP achieves
a great level of consistency between (joint modeling of) tour and stop generation but is lack of
intra-household interactions. CDAP incorporates intra-household interactions but models stops
conditional upon tours. Search for an approach that would combine the best of both continues.
More details will follow in Session 10 in which we will see how stop frequency is intertwined
with stop location choice and accessibility, and is conditional upon the tour mode.
715
Page 73
73
The described structures like IDAP and CDAP, specifically, DaySim and CT-RAMP are the
most frequently applied in practice. Thats why we used them as prototype in our webinar.
However, there are many other advanced structures proposed in academia and some already
being applied in practice. To name just a few, they include CEMDAP, FAMOS, DASH,
TASHA, ALBATROSS, and ADAPTS. These approaches reflect the ongoing research and
improvements in our profession. In particular, we are looking for a better and more realistic
integration between activity generation, scheduling, and location. We want to better understand
and model how people obey time-space constraints and form travel tours. All modelers are also
looking for more consistency of the generated choice for the same person and between person
within the households. Additionally, with the advent of telecommuting and teleshopping
technologies we need a better understanding and modeling of trade-offs between in-home and
out-of-home activities.
716
Page 74
74
We are done with the main part of our webinar and would like to answer your questions. After
that we will provide a short overview of the recent advances beyond the basic structures
discussed so far and summarize the session.
717
Page 75
Summary
Role and placement of DAP model:
Cornerstone and main distinguishing feature of activity based
modeling
First travel related model that generates activities, tours, and trips
for each person and HH
Applied after population synthesis, long-term models of work and
school locations, and car ownership
Applied before tour/trip destination, mode, and TOD choices
75
To summarize todays webinar we hope you have a better understanding of the role and
placement of the DAP model in the activity based modeling system:
DAPs are the cornerstone and main distinguishing feature of activity based modeling;
First travel related model that generates activities, tours, and trips for each person and HH;
Applied after population synthesis, long-term models of work and school locations, and car
ownership; and
Applied before tour/trip destination, mode, and TOD choices.
We have also given two practical examples of approaches widely used in practice IDAP and
CDAP:
Individual DAP (IDAP) generates activities, tours, and trip in a consistent way for each
person independently; and
Coordinated DAP (CDAP) considers interactions between HH members and joint travel
explicitly.
718
Page 76
Next Webinar
Executive and Management Sessions
Executive Perspective
Institutional Topics for Managers
Technical Issues for Managers
Technical Sessions
Activity-Based Model Framework
Population Synthesis and Household Evolution
Accessibility and Treatment of Space
Long-Term and Medium Term Mobility Models
Activity Pattern Generation
Scheduling and Time of Day Choice
Tour and Trip Mode, Intermediate Stop Location
Network Integration
Forecasting, Performance Measures and Software
Activity-Based Modeling: Activity Pattern Generation
February 2
February 23
March 15
April 5
April 26
May 16
June 7
June 28
July 19
August 9
August 30
September 20
76
719
Activity type, where mandatory activities have higher weight than non-mandatory activities;
Activity duration, where longer activities are given higher weight; and
Activity location, where all else equal, the longest trip is considered the primary activity.
The means by which travel behavior data is collected plays a role, because certain ways of
collecting data are better at eliciting activity duration, for example. Thus, the quality of the data
may influence the exact criteria used to determine the primary activity.
What are the units of measurement?
Peter: Since activity pattern is a choice model, it deals with utilities, which is measured in utiles.
Differences in utility are proportional to changes in the probability of selecting a given pattern;
+1 utile means that the choice probability is approximately doubled, while -1 utiles means that
choice probability is approximately halved.
How and when does mode choice figure in an activity based model?
John: This webinar focused on activity pattern generation; that is, the type, sequence and number
of activities performed during the day. Mode choice enters at a later step in the model sequence.
It is a separate module and will be discussed in Webinar #10. In general terms, once the activity
pattern is known, then a mode choice is made for each tour in the day-pattern. But it may be
more involved, because in some models the activity pattern changes after the choice of mode.
For example, after the tour mode (and location and time-of-day) are chosen, there may be a
residual time window in the day-pattern that allows for an intermediate stop in one of the tours.
How do activity-based models account for summer school when predicting school trips?
Peter: Like trip-based models, activity based models model a regular weekday. In this sense,
most ABMs do not model seasonal differences, which may include summer school, vacation, and
season-specific resident populations, among others. These models also do not represent weekend
travel, which can be very different from weekday travel. The first attempt to have a seasonal
component in an activity based model is the Maricopa Association of Governments model. It is
currently under development, so we cannot share any findings yet. The short answer is that the
model needs to be segmented by season. In the case of Phoenix, there are important seasonal
differences in the resident population and in travel patterns related to weather differences. Some
720
regions attempt to incorporate seasonal effects by averaging across seasons. This is incorrect, as
it doesnt represent any real condition. If only one model can be maintained, it is best to choose a
season to model and develop complementary methods to account for seasonal differences.
721
722
Page 1
Activity-Based Modeling
Session 9: Scheduling & Time-of-Day (TOD) Choice
723
Page 2
Acknowledgments
This presentation was prepared through the collaborative efforts
of Resource Systems Group, Inc. and Parsons Brinckerhoff.
Presenters
Peter Vovsha and Maren Outwater
Moderator
Stephen Lawe
Media Production
Bhargava Sana
This is a collective effort of RSG & PB. It is largely built on our experience with many activitybased models in practice.
Peter Vovsha and Maren Outwater are co-presenters. They were also primarily
responsible for preparing the material presented in this session.
Stephen Lawe is the session moderator.
Content development was also provided by Joel Freedman, John Gliebe, and Rosella
Picado. John Bowman provided review.
Bhargava Sana was responsible for media production, including setting up and managing
the webinar presentation.
724
Page 3
February 2
February 23
March 15
April 5
April 26
May 17
June 7
June 28
July 19
August 9
August 30
September 20
725
Page 4
Learning Outcomes
Role and placement of TOD choice in ABM
Advantages of ABM TOD approach with fine
temporal resolution vs. traditional peak factors
Structure of TOD choice model and alternatives in
choice set
Consistency of individual daily schedules with all
activities, trips, and tours w/o gaps or overlaps
Main variables explaining individual TOD choice
TOD choice sensitivity to congestion, pricing, and
other policies
Activity-Based Modeling: Scheduling & TOD
You will learn today about the role and placement of time-of-day choice in the activity-based
model system as well as important advantages of activity-based models with respect to time-ofday choice, in particular a finer level of temporal resolution compared to traditional peak factors
and other simplified techniques. We hope you will have a clear idea how the time-of-day choice
model is structured and how the main alternatives are specified. Another important lesson is to
understand how consistency of individual daily schedules can be ensured when all activities,
tours, and trips are scheduled w/o gaps or overlaps. This is something that is principally different
from the 4-step approach. Finally, you will learn about main factors and corresponding variables
explaining time-of-day choice and ensuring its sensitivity to congestion, pricing, and other
policies.
726
Page 5
Outline
Basic terminology
Temporal level of resolution for different TOD choice
models
Structure of statistical models for TOD choice with fine
temporal resolution
Examples of statistical analysis and model estimation
Individual daily schedule consistency and concept of
dynamically updated time windows
Examples of TOD choice model validation and policy
analysis
Ongoing research, main directions, and challenges
Activity-Based Modeling: Scheduling & TOD
We will start with the basic terminology and then consider such topics as:
727
Page 6
Entire Tour
7:00am
Start (outbound)
7:10am
Stop at Starbucks
7:20am
7:50am
Arrive at work
12:00am
12:50am
Return to workplace
5:00pm
5:30pm
6:40pm
7:20:pm
End (inbound)
8:00pm
Start
Primary Activity
Start
Tour duration
12hours 20min
Activity duration
8hours 10min
End
Tour time-of-day choice is a choice of tour start and end times. It is a 2-dimensional
characteristic. In this example, we have a person who departs from home to work, stops for
breakfast on the way, have a lunch break, then depart from work, visit a shopping mall on the
way home and finally arrives back home. The entire framework from 7am until 7:20pm is
spanned by a single work (commuting) tour that lasts more than 12 hours. This is quite a usual
case, not an extreme one since the tour framework may include multiple activities and trips. If
we want to single the primary activity on this tour, it is work that starts at 7:50 and ends at 5pm.
The duration of the primary activity is 8 hours 10 min., and includes a work-based sub-tour for
lunch.
728
Page 7
Trip
7:00am
Departure
7:10am
Stop at Starbucks
Arrival
7:20am
Departure
7:50am
Arrive at work
Arrival
12:00am
Departure
12:50am
Return to workplace
Arrival
5:00pm
Departure
5:30pm
Arrival
6:40pm
Departure
7:20:pm
Arrival
Duration 40 min
8:00pm
Departure
Activity at Destination
Duration 10min
If we focus on particular trips then for each trip we need to model departure time, arrival time,
and corresponding activity duration at the destination. In most activity-based models tours are
scheduled first and then trip-level details are added conditional upon the tour schedule. In
practical terms only one of dimensions (departure time) has to be modeled, while other
dimensions like arrival time and duration are identified by the schedule information of prior
modeled activities and the travel time required for the trip.
729
Page 8
In addition to formal consistency, we also control for reasonability of travel times and activity
duration. This level of analysis and modeling is not possible at all with an aggregate 4-step
730
model. Day schedule consistency is essential for portraying congestion pricing impacts. If one
trip or activity is rescheduled in reality it would trigger a chain of adjustments for other trips.
731
Page 9
5 min (ABM/trips)
288 bins
3:00am, 3:01am
3:00am-3:04am
3:05am, 3:06am
3:05am-3:09am
3:25am, 3:26am
3:25am-3:29am
5:55am, 5:56am
5:55am-5:59am
5:30am-5:59am
6:00am, 6:01am
6:00am-6:04am
6:00am-6:29am
8:30am, 8:31am
8:30am-8:34am
8:30am-8:59am
4:00am, 4:01am
4:00pm-4:04am
4:00pm-4:29pm
30 min (ABM/tours)
48 bins
Aggregate TOD
periods (4-Step)
3:00am-3:29am
Night
AM
PM
First, we need to distinguish between 4 levels of temporal resolution (or accuracy of the time-ofday choice model) and we have to understand the corresponding implications for the model
structure in terms of number of time-of-day alternatives. The most exact way is to operate with
continuous time that some advanced models do. This means that we literally have 1,440 min to
model explicitly as choices for each trip departure time, etc. The second best is to operate with a
5-min resolution that results in 288 time bins per day. It is enough in many respects for planning
purposes, in particular, for portraying congestion effects. It can be applied in activity-based
models today at least for the trip-level time-of-day. At the tour-level it is more realistic to operate
with a 30-min resolution that results in 48 time bins per day. We will be using this level of
temporal resolution in many illustrations in this webinar. Finally, in aggregate 4-step models we
are forced to use even cruder time-of-day periods that may span several hours. It is always a
good idea to operate with a high level of temporal resolution but it has to be balanced with the
model complexity and associated run time.
732
Page 10
10
Time-of-day choice plays a very important role in an activity-based model system. An essential
feature of an advanced activity-based model is a consistent scheduling of all activities, trips, and
tours for each individual. It is an integral component of an activity-based model and of the daylevel approach in general. It is also yet another major feature differentiating activity-based model
from 4-Step. Activity-based models have principal advantages over 4-step models w.r.t. time-ofday choice such as a fine level of temporal resolution and sensitivity to congestion, pricing, and
transit improvements. We will be again focusing a regular weekday and all statistics that you will
see relate to a regular weekday. In particular, commuting time-of-day patterns and associated
congestion effects will be the primary focus of our discussion.
733
Page 11
Mobility
Choices
Model
Inputs
Long-Term
Choices
Daily Activity
Pattern
Tour TOD
Details
Transport
level-of-service
and
accessibilities
Trip TOD
Details
Network
assignments
Model Outputs
11
Time-of-day choice is not a single mode component. Time-of-day choice is applied for each
travel tour in the package of tour-level models. This will be the primary topic for our webinar.
Subsequently, time-of-day choice is modeled for each trip within each tour in the package of
trip-level models. When we model time-of-day choice for each tour we already know the
outcomes of the upper-level models. In particular we know, the person and HH characteristics,
location of work and school for each worker and student, car ownership and other mobility
attributes, and we know the daily activity pattern for each person in terms of number of tours and
main activities. Now we have to schedule these activities and model the corresponding tour/trip
departure and arrival times.
734
Page 12
12
It is important to recognize that the conventional 4-step structure is very limited with respect to
time-of-day choice. First of all, it is very difficult to find a place of time-of-day choice that
would be the 5th step. Different schemes were tried in the past and none of them were
satisfactory. Secondly, it is difficult to move from the crude time-of-day periods to finer level of
resolution. Thirdly, we lose a tour-level consistency between different trips. For this reason in
many 4-step models some simplified assumption on round-trip symmetry were applied. Finally,
it is difficult to prepare and apply a 4-step model for congestion pricing studies. It results in an
infeasible number of time-of-day slices and segments. The micro-simulation activity-based
model framework that we discuss today offers a much better solution.
735
Page 13
Build Alternative(s)
13
Lets consider a transportation planning and policy project that might be faced by an MPO or
DOT and how daily activity pattern modeling fits into the picture. We have used this example in
several previous sessions to talk about how activity-based modeling components might affect the
analysis.
For this scenario analysis, we will be considering a number of alternatives: a no-build alternative
and a various configurations of the build alternative. In the no-build alternative the bridge has 4
lanes (2 in each direction), there are no tolls, and the transit fare stays the same all day. In the
various build alternatives, there are 6 lanes on the bridge. In some alternatives the two additional
lanes will be HOV lanes all day, while in other alternatives the two additional lanes will be HOV
lanes only during peak periods. In addition, in some build alternatives there will be a new toll
that is the same across the entire day, while in other build alternatives there will be a toll that will
be only applied during peak periods, or when certain levels of congestion occur. Finally, in the
build alternatives regional transit fares will be higher during peak periods.
736
Page 14
14
For this bridge example, DAP generation may be affected in the following ways. If a persons
accessibility is improved by the bridge, the frequency of discretionary activity episodes may
increase. In order to take advantage of HOV lanes, household members may decide to share rides
to work. This not only affects mode choice, but also affects daily activity patterns, because of the
coordination of patterns between household members. It may even lead to increased non-work
joint activities.
737
Page 15
15
To summarize our discussion so far and move to some operational models, there is a general
tendency to improve the level of temporal resolution. Continuous duration models represent the
best solution in this regard but they have their own limitations. We will discuss today some
practical compromises that can be found in most applied activity-based models in practice that
includes such aspects as:
738
Page 16
Resolution
60 min
30 min
5 min
Continuous,
1 min
Model
Trip departure
24
21
Tour TOD
24(24+1)/2 = 300
21(21+1)/2=231
Trip departure
48
40
Tour TOD
48(48+1)/2 = 1,176
40(40+1)/2 = 820
Trip departure
288
230
Tour TOD
288(288+1)/2 =
41,616
230(230+1)/2 =
26,565
Trip departure
1,440
1,142
Tour TOD
1,440(1,440+1)/2 =
1,037,520
1,142(1,142+1)/2 =
652,653
Limits of discrete
technique
16
When we discretize time to apply choice models that are easy to incorporate in practice, we have
to stop at a feasible level of temporal resolution. In the table shown above, the green areas
correspond to feasible choice structures with fewer than 1,000 alternatives. The red areas are
problematic for a discrete choice model. Some reasonable simplifications are applied in practice,
for example, the night time can be collapsed into a single period since we do not have many trips
there. This corresponds to the last column. You can see that both tour and trip time-of-day can be
implemented with a 30-min temporal resolution (that is the prevailing state-of-the practice at the
moment). Trip departure time choice can be implemented with even a finer level of temporal
resolution.
739
Page 17
17
How does it really work and how can we form a choice model and corresponding utility when
choosing a time bin? Do we have to form hundreds or thousands of unique utility expressions?
This is not feasible. The key technical approach is to mimic a continuous duration model in
discrete space by means of so-called shift variables. The main difference between a shift variable
and ordinary variable as you can see on the slide is that a shift variable enters all utility
expressions with the same coefficient but it is additionally multiplied by the time itself. This
time-related multiplier creates a shifting effect that mimics a continuous duration model in
discrete space.
740
Page 18
0.2500
0.2000
Shift
0.1500
Base
0.1000
0.0500
0.0000
6
10
11
12
13
14
15
16
17
18
19
20
21
22
18
This structure allows for a single variable to affect the entire temporal distribution. Lets
consider this distribution of, say, arrival back home from work. Lets say we have a base case as
shown on the slide.
741
Page 19
0.2500
0.2000
Shift
0.1500
Base
0.1000
0.0500
0.0000
6
10
11
12
13
14
15
16
17
18
19
20
21
22
19
If we add a shift variable that is positive (for example high income) it would shift the entire
distribution to later hours.
742
Page 20
0.2500
0.2000
Shift
0.1500
Base
0.1000
0.0500
0.0000
6
10
11
12
13
14
15
16
17
18
19
20
21
22
20
If we add a shift variable that is negative (for example part-time work) it would shift the entire
distribution to earlier hours. This is a very powerful technique in practice that we widely apply.
743
Page 21
U(0)=A(0)+0XB+02XC+03XD
U(1)=A(1)+1XB+12XC+13XD
U(2)=A(2)+2XB+22XC+23XD
.
21
The technique of shift variables has been extended recently to accommodate more elaborate nonlinear effects including polynomial functions, piece-wise functions, trigonometric functions, etc.
In all these case the shift variable and coefficient are multiplied by a certain predetermined
function of time. The product is called temporal profile that is a component of the time-of-day
choice utility function. Temporal profiles are convenient to analyze in graphical form (examples
will be shown).
744
(0hrs)
Before 5 am(0.5hrs)
5:00 am to 5:30 am (1hrs)
5:30 am to 6:00 am (1.5hrs)
6:00 am to 6:30 am(2hrs)
6:30 am to 7:00 am(2.5hrs)
7:00 am to 7:30 am(3hrs)
7:30 am to 8:00 am(3.5hrs)
8:00 am to 8:30 am(4hrs)
8:30 am to 9:00 am(4.5hrs)
9:00 am to 9:30 am(5hrs)
9:30 am to 10:00 am(5.5hrs)
10:00 am to 10:30 am(6hrs)
10:30 am to 11:00 am(6.5hrs)
11:00 am to 11:30 am(7hrs)
11:30 am to 12:00 pm(7.5hrs)
12:00 pm to 12:30 pm(8hrs)
12:30 pm to 1:00 pm(8.5hrs)
1:00 pm to 1:30 pm(9hrs)
1:30 pm to 2:00 pm(9.5hrs)
2:00 pm to 2:30 pm(10hrs)
2:30 pm to 3:00 pm(10.5hrs)
3:00 pm to 3:30 pm(11hrs)
3:30 pm to 4:00 pm(11.5hrs)
4:00 pm to 4:30 pm(12hrs)
4:30 pm to 5:00 pm(12.5hrs)
5:00 pm to 5:30 pm(13hrs)
5:30 pm to 6:00 pm(13.5hrs)
6:00 pm to 6:30 pm(14hrs)
6:30 pm to 7:00 pm(14.5hrs)
7:00 pm to 7:30 pm(15hrs)
7:30 pm to 8:00 pm(15.5hrs)
8:00 pm to 8:30 pm(16hrs)
8:30 pm to 9:00 pm(16.5hrs)
9:00 pm to 9:30 pm(17hrs)
9:30 pm to 10:00 pm(17.5hrs)
10:00 pm to 10:30 pm(18hrs)
10:30 pm to 11:00 pm(18.5hrs)
11:00 pm to 11:30 pm(19hrs)
11:30 pm to 12:00 am(19.5hrs)
After 12:00 am(20hrs)
Page 22
10
Duration- ptworker
Duration- student
-2
-4
-6
-8
22
As we discussed before, impact of shift variables can be illustrated in graphical from. For
example, for part-time workers in the San Diego activity-based model, there is a logical tendency
to avoid very early hours for departure from home (red line) and also a tendency to prefer shorter
durations (violet line). It is quite logical since majority of part-time workers are females,
frequently with children.
745
(0hrs)
Before 5 am(0.5hrs)
5:00 am to 5:30 am (1hrs)
5:30 am to 6:00 am (1.5hrs)
6:00 am to 6:30 am(2hrs)
6:30 am to 7:00 am(2.5hrs)
7:00 am to 7:30 am(3hrs)
7:30 am to 8:00 am(3.5hrs)
8:00 am to 8:30 am(4hrs)
8:30 am to 9:00 am(4.5hrs)
9:00 am to 9:30 am(5hrs)
9:30 am to 10:00 am(5.5hrs)
10:00 am to 10:30 am(6hrs)
10:30 am to 11:00 am(6.5hrs)
11:00 am to 11:30 am(7hrs)
11:30 am to 12:00 pm(7.5hrs)
12:00 pm to 12:30 pm(8hrs)
12:30 pm to 1:00 pm(8.5hrs)
1:00 pm to 1:30 pm(9hrs)
1:30 pm to 2:00 pm(9.5hrs)
2:00 pm to 2:30 pm(10hrs)
2:30 pm to 3:00 pm(10.5hrs)
3:00 pm to 3:30 pm(11hrs)
3:30 pm to 4:00 pm(11.5hrs)
4:00 pm to 4:30 pm(12hrs)
4:30 pm to 5:00 pm(12.5hrs)
5:00 pm to 5:30 pm(13hrs)
5:30 pm to 6:00 pm(13.5hrs)
6:00 pm to 6:30 pm(14hrs)
6:30 pm to 7:00 pm(14.5hrs)
7:00 pm to 7:30 pm(15hrs)
7:30 pm to 8:00 pm(15.5hrs)
8:00 pm to 8:30 pm(16hrs)
8:30 pm to 9:00 pm(16.5hrs)
9:00 pm to 9:30 pm(17hrs)
9:30 pm to 10:00 pm(17.5hrs)
10:00 pm to 10:30 pm(18hrs)
10:30 pm to 11:00 pm(18.5hrs)
11:00 pm to 11:30 pm(19hrs)
11:30 pm to 12:00 am(19.5hrs)
After 12:00 am(20hrs)
Page 23
1.5
Departure -Income (<=30K)
1
Departure -Income (30K to 60K)
0.5
Duration- (>=100K)
-0.5
-1
-1.5
23
These are examples of income effects from the same activity-based model. It can be seen for
example, that low-income workers (less than $30K, blue line with squares) tend to start much
earlier compared to medium-income workers that serve as the base case. Interesting duration
effect can be observed for medium-income workers (blue line with crosses). They most
frequently prefer normal fixed-schedule workday (9-10 hours including commuting) compared to
say high-income workers (orange line) show have more flexible schedules.
746
Page 24
-0.1000
-0.2000
-0.3000
Departure -Female w/ Preschooler
Departure-Female
-0.4000
-0.5000
-0.6000
-0.7000
-0.8000
-0.9000
-1.0000
24
Another set of shifts correspond to gender effects. For example, you can see that arrival time
back for female workers with a preschool child is (red line) highly concentrated around a
relatively early hour (4:30pm). There are many other different shift effects incorporated in either
departure, or arrival, or duration components.
747
Page 25
25
Temporal profiles modeled for each travel purpose and person type as a combination of multiple
impacts and shifts. They are compared to the observed distributions across multiple dimensions
at the validations stage. This will be discussed further in Part 2 of this session, with multiple
examples.
748
Page 26
26
By applying continuous shift models in discrete space we mimic the good properties of
continuous models such as variety of variables and corresponding profiles as well as quite a
parsimonious structure with a few parameters to estimate. The statistical model structure is
actually very simple. It is an ordinary choice model, most frequently logit. You can estimate it
using a standard software package that you would use to estimate a mode choice model. The
number of coefficients to estimate is less than the number of alternatives, thanks to the shift
variables. However continuous time models have their own merits, such as better and more
natural incorporation of activity duration as well as possible integration with discrete choice
models.
749
Page 27
Practically
TOD choice integrated with SUE
by 6-12 TOD periods (carrying
over incomplete trips from period
to period)
Trip tables and LOS variables
aggregated by 6-12 TOD periods
HH, person, and zonal variables
differentiate beyond TOD periods
Activity-Based Modeling: Scheduling & TOD
AM peak (7am-9am)
Late AM shoulder (9am-10am)
Midday (10am-3pm)
Early PM shoulder (3pm-4pm)
PM peak (4pm-6pm)
27
While the time-of-day choice operates with temporal resolution of 30 min or less we still have
limitations on the network assignment side. It would be difficult to run 40 half-hour static
assignments to match the activity-based model resolution. An ideal solution would be a full day
DTA but this is still problematic for large regions. The compromise solution is to apply 5-8 static
assignments that portray the main differences in time-of-day periods with respect to congestion
and pricing as in the examples shown on the slide. These assignments can be run in parallel with
distributed processing so that the run time will be equal to the single assignment run time.
Distributed processing is possible with multiple computers or with multiple cores and threads on
a single server.
750
Page 28
28
Now lets consider details of the time-of-day choice model structure. We will focus on a single
travel tour (for example work tour). We consider entire travel tour as unit of modeling. The
model represents joint choice of:
751
Page 29
29
In this choice structure we have 820 choice alternatives that correspond to feasible combinations
of the departure half-hour bins and arrival half-hour bins. Arrival back home cannot be earlier
than departure from home. However we do not have to estimate and apply 820 unique utility
expressions. That would be impossible. The advantage of this structure is that it can be
decomposed into 120 meaningful dimensions for which we have to form utility components
using shift variables. These components are further combined for each of the 820 alternatives.
The first set of utility components corresponds to 40 departure time alternatives. The second set
of utility components corresponds to 40 arrival time alternatives. The third set of utility
components corresponds to 40 possible tour/activity durations. These combinations can be
illustrated in the following way.
752
Page 30
24
30
Consider a work tour that we want to predict for a given person. Currently the entire day window
is open for the person and no other activities have been scheduled yet.
753
Page 31
24
31
There are many factors and variables that affect the departure time from home for a work tour.
These include office hours, the desire to avoid congestion and the potential need to provide rides
to children. They are all included in the corresponding utility components.
754
Page 32
Work tour
10
24
32
Lets say that based on the departure time utility component only the optimal time for this person
to start work is 10am. This is the best utility so far.
755
Page 33
Work tour
10
24
33
There are also many factors and variables affecting time of arrival back home. These include
(again) office hours and the desire to avoid congestion, but in the particular example the
individual may wan t to play tennis before it gets dark. These factors are incorporated in the
second utility component for each alternative. Lets say that from the perspective of arrival time
back home we can find an optimal solution (i.e., the best utility).
756
Page 34
Work tour
10
15
24
34
And this optimal solution will look like this. So far there is nothing in the choice structure that
would prevent the choice model form generation a solution like this. Interestingly, this is not just
a theoretical absurd. A 4-step trip-based framework where a time-of-day choice model is applied
for each trip separately can easily generate a solution like this. However, in the tour-based
framework we have a third set of factors and variables encapsulated in the duration component
of the utility function (push button). These considerations would results in a more realistic
solution where all three dimensions departure, arrival, and duration would be integrated and the
best compromise will be found.
One contrast between a 4-step model and an activity-based model would be their response to
congestion pricing. A 4-step model in which outbound and inbound commuting trips are
considered separately can shift the morning commuters to a later period and evening commuters
to the earlier period, depending on the toll structure and schedule. However, this response is
illogical since the overall balance between outbound and inbound time should be kept. This can
only be achieved with an activity-based model that has an activity duration dimension explicitly.
757
758
Page 35
Work tour
19
24
35
This solution will look most probably like this, i.e. what we most frequently observe in reality. In
this example we assume a 10-hour entire-tour span because it includes commuting time in
addition to a normal 8-hour workday.
759
Page 36
36
Similar logic is used in many other activity-based models. For example in the DaySim activitybased model the dimensions of tour are redefined in the following way:
Joint choice of arrival time at primary destination and departure time from primary
destination is modeled first.
The entire tours duration, departure time from home, and arrival time back home are
modeled later after stops are added.
The number of alternatives is slightly less compared to CT-RAMP since the tour framework is
limited to the primary activity only.
760
761
Page 37
Arrive
at
Work
Depart
from
Work
Arrive
at
Shop
7:30am
8:00am
5:00pm
5:20pm
Depart
from
Shop
6:30pm
Arrive
back
Home
7:00pm
DaySim approach:
Focus on primary
activity
Entire-tour details
added outward
CT-RAMP approach:
Start with entire tour framework (convenient for constructing day schedule)
Tour details added inward by inserting stops and departure times
37
To summarize differences between two approaches lets consider a realistic tour structure as
shown on the slide:
Inward tour window partition (CT-RAMP) is characterized by the following main features:
Tour time-of-day modeled from departure from home until arrival back home; entire tour
window constrains trip departure time
762
Stops are sequentially inserted in a chronological order for each half-tour (outbound,
inbound)
Outward tour window extension (DaySim) is characterized by the following main features:
Primary tour activity time-of-day modeled from arrival at primary destination until
departure from primary destination; primary activity window constrains trip departure
time
Stops are sequentially added in a chronological order for each half-tour (outbound,
inbound)
Both approaches have their own merits and eventually provide all necessary tour details.
763
Page 38
Simplified Example
Commuting tours to work
1 hour temporal resolution (instead of 30 min)
Complete prototype TOD structure but the choice set
is limited to a subset of most frequent alternatives
Real stats from Bay Area Travel Survey (BATS), 2000
38
For you to have a better feel and more technical hand-on details we consider a simplified
example of commuting tour. We will model it with a 1-hour resolution to reduce the number of
alternatives. We will consider a complete prototype choice model structure but to further limit
the choice set we will consider only a subset of most frequent alternatives. WE define them
based on the real stats from the BATS survey.
764
Page 39
Modeled
Modeled
25%
20%
15%
10%
5%
0%
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
Departure
Activity-Based Modeling: Scheduling & TOD
Arrival
39
The modeled areas cover majority of cases. They correspond to 5 morning hours for departure
time and 5 afternoon hours for arrival time back home.
765
Page 40
3-5
6-10
11-15
16-20
21-27
3-5
0.0%
0.1%
1.2%
3.4%
0.3%
0.6%
9.3%
62.9%
8.0%
1.0%
6.1%
2.7%
0.9%
2.0%
6-10
11-15
16-20
21-27
0.3%
40
If we single out tours that start in the first selected period and end in the second selected period
we cover more than 60% of the observed tours. In the real time-of-day choice application we of
course consider all possible combinations. But for now we will focus on the selected time
periods for simplicity (and only to reduce the number of alternatives).
766
Page 41
Departure from home
6 (5:30-6:29 AM)
16 (15:30-16:29 PM)
10
17 (16:30-16:29 PM)
11
18 (17:30-16:29 PM)
12
19 (18:30-16:29 PM)
13
20 (19:30-16:29 PM)
14
99
17 (16:30-16:29 PM)
10
18 (17:30-16:29 PM)
11
19 (18:30-16:29 PM)
12
20 (19:30-16:29 PM)
13
10
16 (15:30-16:29 PM)
11
17 (16:30-16:29 PM)
12
18 (17:30-16:29 PM)
10
13
19 (18:30-16:29 PM)
11
14
20 (19:30-16:29 PM)
12
15
16 (15:30-16:29 PM)
16
17 (16:30-16:29 PM)
17
18 (17:30-16:29 PM)
18
19 (18:30-16:29 PM)
10
19
20 (19:30-16:29 PM)
11
20
16 (15:30-16:29 PM)
21
17 (16:30-16:29 PM)
22
18 (17:30-16:29 PM)
23
19 (18:30-16:29 PM)
24
25
Duration
Alternative
Utility
8 (7:30-8:29 AM)
9 (8:30-9:29 AM)
10 (9:30-10:29 AM)
Activity-Based Modeling:
Scheduling & TOD
20 (19:30-16:29 PM)
10
DEP7DEP7
+ ARR16
DUR9
+ +ARR16
+ DUR9
41
This results in 25 alternatives, as listed in the table. The choice set includes all combinations of 5
possible departure times and 5 possible arrival times. For each of the alternatives we have three
utility components that describe the corresponding departure time, arrival time, and duration.
Consider for example alternative number 6 (push the button). This alternative assumes departure
from home at 7am, arrival back home at 4pm and total tour duration of 9 hours.
767
Page 42
42
A model of this type is estimated based on the conventional Household Travel Survey It has to
be processed in the tour format and reported travel time rounded to the nearest half-hour (bin).
LOS variables and mode choice log-sums are specified by broader time-of-day Periods.
No sampling needed, all 820 alternatives are modeled. Parsimonious utility structure is applied
with 35-40 constants and 30-55 other coefficients. Despite a limited number of parameters
statistical fit is much better than for the reference model with 820 constants because of the shift
variables that capture many impacts.
768
Page 43
43
769
Page 44
44
Models were internally validated against observed departure, duration and arrival patterns across
different segmentations of the data. Strong effects were found related to:
Most of the estimated effects are very similar for the data sets from Columbus, Atlanta,
Sacramento, San-Diego, Bay Area, and others.
770
Page 45
45
These are some of the stat findings with respect to impact of person and HH characteristics. They
include the following main factors:
771
Page 46
46
As we discussed before, impact of shift variables can be illustrated in graphical from. For
example, work tour arrival times are similar for full-time and part-time workers, but student
workers tend to start work later in the day. This example is derived from the 3-county activitybased model in San Joaquin, Stanislaus and Merced counties.
772
Page 47
47
These are examples of activity duration by purpose from a 3-county activity-based model in the
San Joaquin Valley. Work activities have the longest durations. The distribution peaks around 9
hours, and school activities are second, with peaks around 7 hours. Other activities and workbased activities are quite short in duration, peaking around 30 minutes.
773
Page 48
48
Arrival times by purpose follow expectations for different types of activities, as shown here in
the arrival times by purpose from a 3-county activity-based model in the San Joaquin Valley.
School activities have the highest peak, between 7-8am, with very few arrival times outside this
window. Work activities are spread between 3-9am, with the majority occurring between 6-9am.
Work-based activities are primarily in the middle of the day, between 10am and 4pm. Other
activities tend to be in the evening, between 4-11pm.
774
Page 49
49
In addition to person and HH related effects time-of-day choice is strongly affected by location
and accessibility effects. For example:
Longer travel time in general Extends duration of work tour, Shifts departure from home
to earlier hour, and Shifts arrival back home to later hour;
There is a logical congestion effect: higher travel time impedance in peak periods shifts
trips to and from work to other hours;
Stops on the way to or from the destination extend the tour duration in both directions
(except for escort stops);
Tours to CBD tend to be of longer duration and later in the day (occupation effect); and
Work tours that include sub-tours are of longer duration.
775
Page 50
50
Additionally, the DAP of the person affects the time-of-day choice for each particular tour due to
time-space constraints. In particular, the more tours to schedule in the day, the shorter the
duration of each tour. Higher number of tours tends to shift work and school tours earlier, other
tours later. People generally tend to schedule tours shortly after previous tours to leave a larger
amount of continuous free time for later. All these effects are formalized trough the
corresponding shift variables.
776
Page 51
51
There are also certain effects pertinent to non-work tours. For school tours, there are very
different time-of-day patterns for full- and part-time workers, and for students at various levels
of school. Children stay at school longer when all adults in the household are working.
For shopping, maintenance, and discretionary tours, the likelihood of staying out late in the
evening varies a great deal by age group. Shopping and maintenance tours tend to be of short
duration and restricted to retail hours. Maintenance and discretionary tours implemented jointly
by several household members tend to be longer, relative to those implemented independently.
Again, all these effects are also formalized through the corresponding shift variables.
777
Page 52
52
The micro-simulation activity-based model framework allows for tracking each person through
time to ensure consistency of the generated individual schedule. This is one of the principal
advantages of and activity-based model over 4-step. This feature means basic daily schedule
consistency for each person that includes the following requirements:
These requirements are essential for evaluation of congestion and pricing effects that can be
outside the congestion pricing period. In advanced activity-based models, some additional
features were introduced:
Residual time windows used also for generation of lower-priority activities and tours
(time-of-day intertwined with DAP);
778
779
Page 53
53
An advanced activity-based model like CT-RAMP models joint activities and travel explicitly.
When we model joint tour made by several HH members we have take into account the
following factors (read the first bullet). For a fully joint tour that involves several HH
members we have to ensure that they are all available at the same time. In other words, their
available time windows should have enough of overlap to implement the activity and associated
travel.
780
Page 54
7:00
16:00
3-hour overlap
10:00
19:00
TOD
7:00
22:00
54
Consider example of a couple with the following fixed work schedules. He is only available after
4pm. She is available a couple hours in the morning and after 7pm in the evening. Essentially,
they have only 3 hours to start a joint out-of-home activity. The probability for a joint nonmandatory activity to happen is basically proportional to the residual time window overlap.
781
Page 55
Workers / Non-workers
Work
University / School
University
Work
Maintenance joint
Shopping joint
Discretionary joint
Eating-out joint
Escorting
Shopping individual
Maintenance individual
10
Discretionary individual
11
Eating-out individual
55
All tours including individual and joint are scheduled sequentially in a consistent way where
each subsequent tour can be only scheduled in the residual time window left for this person after
the higher priority tours have been scheduled. Mandatory activities scheduled first, followed by
non-mandatory joint tours, and finally by non-mandatory individual tours. There are some
variations in this order and rules from model to model. However, the basic idea is the same.
782
Page 56
3-Shopping individual
2-Discretionary joint
1-Work
23
56
To illustrate this process of sequential scheduling and technique of residual windows, lets
consider example of a person who plans three activities and tours on the given day.
783
Page 57
3-Shopping individual
2-Discretionary joint
1-Work
7-17
23
57
The work tour is scheduled first w/o scheduling constraints. Lets say it is a conventional
schedule where the worker would leave home at 7 am and would arrive back home at 5 pm.
784
Page 58
3-Shopping individual
1-Work
2-Discret
7-17
20-23
23
58
Now the second tour can go only into the residual time window. Lets assume that this person
scheduled a late discretionary activity like going to a theater between 8 pm and 11 pm.
785
Page 59
1-Work
3-Sh
2-Discret
7-17
18-19
20-23
23
59
Now the third activity must go into the narrow residual window of two hours. Thus it cannot be a
major shopping or distant destination. It is important to recognize and model these
interdependencies because they create many scheduling constraints. Imagine how nave and
unrealistic would be a model that schedules each tour independently.
786
Page 60
Primary destination
Primary destination
Tour mode
Tour mode
Residual window
Stop arrival-departure
Trip mode
Trip mode
60
Now consider how the time-of-day choice model integrated in the activity-based model system
with the other models. Consider first a DaySim type of activity-based model where each person
is modeled separately. For each person, tours are generated by priority and each tour already has
secondary stops generated by the IDAP procedure that we discussed in the previous webinar. For
each tour, the following sequence of sub-models shown in the first column is applied. After the
first tour has been processed, residual time windows are calculated and used as constraints for
scheduling the second tour, etc. An interesting feature of DaySim is that tour time-of-day choice
is applied twice. First, a preliminary time-of-day choice is applied to identify which LOS
variables should be used for destination and mode choice. Secondly, a final tour time-of-day
choice is applied conditional upon the chosen primary destination and tour mode.
787
Page 61
Person 2: Discretionary
tour
Residual
window
overlap
Primary Destination
Tour TOD
Joint tour
Primary Destination
Tour mode
Primary destination
Tour TOD
Tour TOD
Tour mode
Trip mode
Tour mode
Trip mode
Trip mode
61
In the CT-RAMP system, the entire household is considered and the mandatory activities for
workers and students modeled and scheduled first. For person 1 who has a work tour, the
following sub-models shown in the left-most column will be applied. Person 2 does not have a
mandatory activity. The next step involves scheduling of a joint tour that is conditional upon he
residual time window overlap between persons 1 and 2. Finally, individual non-mandatory
activity for person 2 is scheduled conditional upon the residual window left for person 2 after
scheduling the joint tour.
788
Page 62
Highlights
Remarkably good match for Work and School tours with
higher scheduling priority
Reasonable match for Shopping, Maintenance, and
Discretionary activities with lower scheduling priority
Either no or very minor calibration is required
Activity-Based Modeling: Scheduling & TOD
62
How can we validate the time-of-day choice model and prove this concept in practice? During
the model validation process, the activity-based model is applied with time-of-day choice,
aggregate outcomes are compared to an expanded household survey, and, ideally compared to
hourly traffic counts.
This usually results in a very good match to work and school tour times, with higher schedule
priorities and more regularity. Matches to shopping, maintenance and other discretionary
purposes are usually reasonable, but that is because these have lower scheduling priority. Often
calibration of time-of-day choice models is necessary only if you are validating against traffic
counts, since the adjustments to match counts will be needed to allow for differences between the
observed data sources. Lets see some examples from the activity-based models applied in
practice.
789
Page 63
63
This is an example from the San Diego activity-based model for work tour departure-from-home
and arrival-back-home stats. You would probably have a hard time to distinguish between the
observed and modeled time-of-day choice. These distributions are typical. AM peak is a bit
sharper than PM peak, and the model captures all these details quite accurately.
790
1-Before 5 am
2- 5:00 am to 5:30 am
9- 8:30 am to 9:00 am
8- 8:00 am to 8:30 am
7- 7:30 am to 8:00 am
6- 7:00 am to 7:30 am
5- 6:30 am to 7:00 am
4- 6:00 am to 6:30 am
3- 5:30 am to 6:00 am
Page 64
25%
20%
15%
10%
5%
0%
64
The same for school tours. School tours in the Chicago activity-based model are characterized
but very sharp peaks in the morning departure from home and in the evening for arrival back
home.
791
Page 65
65
For shopping tours that is an example of a non-mandatory activity, we also have a good match
but everything gets a bit fuzzier.
792
Before 5 am
11:30 pm to 12:00 am
11:00 pm to 11:30 pm
10:30 pm to 11:00 pm
10:00 pm to 10:30 pm
9:30 pm to 10:00 pm
9:00 pm to 9:30 pm
8:30 pm to 9:00 pm
8:00 pm to 8:30 pm
7:30 pm to 8:00 pm
7:00 pm to 7:30 pm
6:30 pm to 7:00 pm
6:00 pm to 6:30 pm
5:30 pm to 6:00 pm
5:00 pm to 5:30 pm
4:30 pm to 5:00 pm
4:00 pm to 4:30 pm
3:30 pm to 4:00 pm
3:00 pm to 3:30 pm
2:30 pm to 3:00 pm
2:00 pm to 2:30 pm
1:30 pm to 2:00 pm
1:00 pm to 1:30 pm
12:30 pm to 1:00 pm
12:00 pm to 12:30 pm
11:30 am to 12:00 pm
11:00 am to 11:30 am
10:30 am to 11:00 am
10:00 am to 10:30 am
9:30 am to 10:00 am
9:00 am to 9:30 am
8:30 am to 9:00 am
8:00 am to 8:30 am
7:30 am to 8:00 am
7:00 am to 7:30 am
6:30 am to 7:00 am
6:00 am to 6:30 am
5:30 am to 6:00 am
5:00 am to 5:30 am
Page 66
9%
8%
7%
6%
5%
4%
3%
2%
1%
0%
66
The validation results are also very reasonable for non-mandatory purposes, but not as perfect as
for work and school. This is an example of the validation for discretionary tours with the SanDiego activity-based model.
793
Page 67
67
Another important dimension is the tour duration. Again the San Diego model replicates the
observed pattern almost exactly. The distribution is typical and looks similar in many other
metropolitan regions. The mode duration is about 10.5 hours because is includes the entire tour
and not only the work activity itself.
794
Page 68
68
Shopping tour duration distribution from the San Diego activity-based model is also replicated
very well. The average duration of shopping tour is of course much shorter (1.5-2 hours) and the
distribution is much sharper compared to work tours.
795
Page 69
69
Why it is better for Work and School? There are several reasons for that:
Work and school activities have clear schedules and it is easier to relate them to person
characteristics;
Work and school tours are modeled first in the scheduling chain; non-work activities are
subject to compounding of small errors; and
Improvements in entire-schedule conditionality and sequence of scheduling steps are on
the way.
796
Page 70
70
This additional level of validation is critical for time-of-day choice models, since there is likely
to be a discrepancy between the traffic counts and the household survey at the half-hour or hour
time periods. Household surveys can be lumpy, given humans tendency to round times to the
nearest 5 or 15 minute increments. Also, expansion of household surveys can be lumpy at the
specific origin and destination level. Traffic counts will include all vehicles, and at a minimum,
trucks should be excluded so that the time-of-day choice model can be validated against autos.
This will still include some commercial vehicles that are autos or light trucks, but the majority of
the auto volumes will be for passenger movements.
797
Page 71
We present results
(Global) Absolute number of toll users vs. the base
(Congestion) Absolute number of toll users vs. the base
(Congestion) TOD distribution of toll users vs. the base
(peak spreading effect)
71
We would like to present the time-of-day choice model performance for pricing policies. These
evaluations were implemented with the Chicago activity-based model, 2011. Two pricing
scenarios were evaluated and compared to the base scenario:
A global pricing scenario in which current tolls are multiplied by a factor of five on all
toll facilities during the entire day; and
A congestion-based pricing scenario in which the tolls are increased by five times for
only the 2-hour AM and PM peak periods.
For the global pricing scenario, we consider differences in the number of toll users; and
For the congestion pricing scenario, we consider differences in the number of toll users
overall, and by time period.
798
Page 72
72
Global pricing with the radical rise of all tolls logically resulted in a significant reduction of toll
users in the Chicago activity-based model. Only a few users with very high VOT continue using
toll facilities. Please note also, that the existing toll users primarily use toll facilities in peak
periods when the parallel facilities are congested.
799
Page 73
73
The congestion pricing scenario in the Chicago activity-based model yielded a very different
outcome with the number of users affected primarily in the peak periods when tolls were raised.
The number of toll users outside the peak periods barely changed.
800
Page 74
Peak-Spreading Effect
74
If we compare temporal distribution of toll users in the Chicago activity-based model before and
after rather than absolute numbers we can see an interesting peak-spreading effect where the
morning peak was slightly shifted to earlier hours while the evening peak was slightly shifted to
later hours. This is a consequence of the fact that high income (high VOT) users that stayed on
the toll road after the tolls have been raised are characterized by a relatively longer work duration
compared to low-income users who switched to non-toll roads or transit.
801
Page 75
Adjusted
6
5
4
3
2
1
15.00
14.00
12.00
13.00
11.00
10.00
8.00
9.00
7.00
6.00
4.00
5.00
3.00
2.00
1.00
Duration
Target: Fulltime
Workers
Original
% of Tours
Fewer individual
work activities
Longer individual
work durations
Aggregate work
durations constant
75
This is a travel demand management scenario for the Burlington (VT) activity-based model that
was conducted as part of the SHRP 2 C10 research. This scenario demonstrates that when people
are working under flexible schedules, then work durations are longer for full time workers.
802
Page 76
30000
20000
10000
03:00
04:00
05:00
06:00
07:00
08:00
09:00
10:00
11:00
12:00
13:00
14:00
15:00
16:00
17:00
18:00
19:00
20:00
21:00
22:00
23:00
00:00
01:00
02:00
76
In addition, the Burlington activity-based model shows a reduction in peak period and midday
travel, since flexible schedules often include 1-2 days off every 2 weeks. And, the start times are
earlier as well as the time to return home from work in the evening.
803
Page 77
30000
BASE-NONWORK
TDM-NONWORK
20000
10000
03:00
04:00
05:00
06:00
07:00
08:00
09:00
10:00
11:00
12:00
13:00
14:00
15:00
16:00
17:00
18:00
19:00
20:00
21:00
22:00
23:00
00:00
01:00
02:00
77
This also translates into fewer overall work trips and allows for more non-work trips in the
Burlington activity-based model, which are conducted on days off or in the morning or
evening before or after work. There are fewer trips in the mid-day most likely because workers
are going to work on fewer days over a two-week period.
804
Page 78
78
Time-of-day choice and related activity scheduling is a very dynamic filed with many research
directions pursued by different researches. To name just a few promising directions, there is
ongoing development in terms of flexible correlation patterns that account for differential
similarities across time-of-day alternatives, many suggestions to enrich the functional form of the
utility, as well as many interesting attempts to integrate time-of-day choice with other choice
dimensions such as mode, destination, car allocation, etc.
805
Page 79
79
There are also many alternative ways to construct daily schedules and move towards a
continuous representation of time, integration of activity-based model and DTA as part of the
SHRP 2 program, integrated activity generation and scheduling (or time allocation) procedures,
etc. Some researches go beyond the daily framework and consider entire week for modeling
individual schedules. Some other researches pursue a multi-stage scheduling procedure with replanning and consolidation rules.
806
Page 80
80
Lets take a look at and extended time-of-day choice framework. The ALBATROSS model
provides a good example. It is based on a fundamental behavioral observation that people do not
schedule and implement activities in one day. Some activities (special events) are scheduled
many days in advance and come into daily schedule as pre-fixed. Some activities (shopping)
occur periodically and can be shifted between days, whereas some activities (work, school) occur
daily. Consequently, modeling schedules requires longer time horizon (at least one week) and
follows a set of steps:
807
Page 81
81
In summary we would like to mention the following key points. Time-of-day choice is a key
component of activity-based modeling. It is closely intertwined with tour generation, destination
choice, and mode choice. The temporal resolution of these models is improving, from aggregate
time-of-day periods to 30 min and eventually to continuous time.
Tour-level time-of-day choice is actually the joint choice of departure from home (or arrival at
primary destination) and arrival back home (or departure from primary destination). These
choices imply tour duration (or activity duration). Trip-level time-of-day choice is conditional
upon tour time-of-day and includes trip departure time.
The time-of-day modeling framework incorporates a wide variety of variables and effects that
generate consistent individual daily schedules and are realistically sensitive to congestion and
pricing. Such models have been successfully applied in many activity-based models in practice
and tested for many policies.
808
Page 82
82
809
Page 83
83
810
Page 84
Next Webinar
Executive and Management Sessions
Executive Perspective
Institutional Topics for Managers
Technical Issues for Managers
Technical Sessions
Activity-Based Model Framework
Population Synthesis and Household Evolution
Accessibility and Treatment of Space
Long-Term and Medium Term Mobility Models
Activity Pattern Generation
Scheduling and Time of Day Choice
Tour and Trip Mode, Intermediate Stop Location
Network Integration
Forecasting, Performance Measures and Software
February 2
February 23
March 15
April 5
April 26
May 17
June 7
June 28
July 19
August 9
August 30
September 20
84
811
scenario. That can be used as a starting point to influence time-of-day choice. That helps get an
approximate time-of-day to start with. Once initial time-of-day is chosen, you can calculate
mode and then actual time-of-day.
Regarding shifts, wouldn't one possible response be to shift a trip from beginning of day to end
of day? How is that reflected in these paradigms?
Peter: Shifts can occur across entire range. Depending on structure, there may be longer shifts.
This depends on the entire structure and other constraints. For example, a person with a very
fixed schedule for work can implement shopping before or after work, since window for work
activity is blocked. This is one of things possible to model with an activity-based model. You
don't just see mode choice or time-of-day change. You see the change of entire day patterns.
In households with multiple children around the same age, do you model their time-of-day
choice together or separately?
Peter: In both CT-RAMP and DaySim, every person is simulated, including small children. Each
person has an individual record and activity schedule. This is applied to 70 year-olds and small
children; however, patterns are different because young children cannot drive and are strongly
linked to the activities of adults.
Is there any treatment, for example iteration, in activity-based models to make sure input hourly
traffic costs are consistent with output hourly traffic costs?
Maren: Yes, the reason we do different iterations is to ensure at the end of the day there is some
convergence. The treatment is something the modeler needs to do. It's not a built-in function, it's
an idea that the iteration must converge and be checked.
Map-21 requires performance measures as a part of project evaluation. Would an activity-based
model be able to do this better than traditional models?
Maren: activity-based models are more detailed spatially and temporally. We also have more
detailed person characteristics. We can then aggregate better for performance measures that are
segmented for different segments of the populations, for example, high- or low-income. This
allows you to get better information about influence of your policies. A lot of people want to
know if a particular policy is going to affect low-income households, for example, and that's
possible with an activity-based model and would be much more difficult with a traditional
model. Having said that, the performance measures that most MPOs are considering have started
to expand to measures that are possible with activity-based models, but they still rely on the
older performance measures, which activity-based models can still handle.
Peter: One of the new measures being extensively discussed is travel time reliability. There is a
growing understanding that travel time reliability has a large impact on people's choices, and
activity-based models can measure reliability very well.
813
814
815
Page 1
Activity-Based Modeling
Session 10: Tour Mode, Primary Destination,
Intermediate Stop Location and Trip Mode
August 9, 2012
816
Page 2
Acknowledgments
This presentation was prepared through the collaborative efforts of Resource
Systems Group, Inc. and Parsons Brinckerhoff.
Presenters
Joel Freedman and John Gliebe
Moderator
Maren Outwater
Media Production
Sumit Bindra, Bhargava Sana
Resource Systems Group and Parsons Brinckerhoff have developed these webinars
collaboratively, and we will be presenting each webinar together. Here is a list of the persons
involved in producing todays session.
Joel Freedman and John Gliebe are co-presenters. They were also primarily responsible
for preparing the material presented in this session.
Stephen Lawe is the session moderator.
Content development was also provided by Jason Chen, Rosella Picado, and Greg
Erhardt. Content review was provided by John Bowman.
Sumit Bindra and Bhargava Sana were responsible for media production, including
setting up and managing the webinar presentation.
817
818
Page 3
February 2
February 23
March 15
April 5
April 26
May 16
June 7
June 28
July 19
August 9
August 30
September 20
3
For your reference, here is a list of all of the webinars topics and dates that have been planned.
As you can see, we have been presenting a different webinar every three weeks. Three weeks
ago, we covered the ninth webinar in the seriesscheduling and time of day choice.
Todays session is the seventh of nine technical webinars, where we will cover the details of tour
and trip modes and intermediate stop location choices. In three weeks, we will discuss how
activity-based modeling demand systems are integrated with network supply models.
819
Page 4
Learning Outcomes
By the end of this session, you will be able to:
Define tour mode
Define trip mode
Explain the importance of consistency between:
Tour mode and trip mode
Tour anchor location, primary destination and stop location
Tour mode and intermediate stop location
Shown on this Page are the webinar learning outcomes. By the end of this session, you will be
able to:
820
Page 5
Webinar Outline
Tour mode, primary destination, intermediate stop
location and trip mode review
Tour mode choice
Intermediate stop location choice
Trip mode choice
Questions and answers
In this webinar, we will cover many of the components of activity-based models that predict
tour-level characteristics of tours, including the location of out-of-home activities on the tour,
and the mode or modes used for the tour. Recall from previous webinars that we distinguish
between primary destination on the tour, and intermediate stops on the tour. The primary
destination on the tour is the location of the primary activity on the tour, or the main reason for
making the tour, whereas intermediate stops are activities that occur between the tour anchor
location (home or work) and the tour primary location. Also note that we distinguish between
tour mode and trip mode. This concept will be explained more fully in a moment.
We will provide opportunities for questions and answers.
821
Page 6
Terminology
Tour mode
Preferred mode or primary mode for the tour
Ensures consistency between modes for each trip on tour
Trip mode
The mode for each trip on the tour
Rubber-banding
The use of out-of-direction distance, time, and/or utility to
choose intermediate stop location
Ensures reasonable locations of stops on tours
There are three terms that we will use in this session that may be familiar, but are worth defining
explicitly, so their meaning is clear.
The tour mode is the preferred mode for a series of trips that begin and end at an anchor location
(typically either home or work). Note that tour mode is not reported in survey data nor is it
directly observed. Rather, it is inferred from the combination of modes used for trips within a
tour. It is defined to ensure a reasonable consistency between the modes used for individual trips
on the tour, as we shall see.
The trip mode is the mode for an individual movement from an origin to a destination. The trip
mode is often, but not always the same as the tour mode that the trip is a part of. It can differ, for
example, if someone carpools to work but takes transit home. It can also differ when someone
drops-off a passenger, going from a shared-ride mode to a drive-alone mode.
Rubber-banding refers to the method of measuring the impedance to an intermediate stop, which
takes into account out-of-direction distance, time, and/or utility to choose an intermediate stop
822
location. In rubber-banding, you can visualize stretching a rubber band between the tours anchor
location and primary destination, holding those two locations fixed. One side of the rubber band
is stretched to consider alternate intermediate stop locations. When this is done, the impedance is
calculated as:
823
Page 7
Key Concepts
A key difference between an ABM and a trip-based model is that in an ABM, each trip is no
longer an independent unit of analysisit is linked with other trips to form tours. The choices
made for one trip affect the outcomes of all other trips on the tour. By accounting for this
dependency among trips (trip-chaining), an ABM can capture more of the factors that truly drive
travel choices. To be effective, however, the model has to ensure consistency among trips on a
tour, and consistency among the different model components. This is a natural extension of good
practice in trip-based modeling, where it is important to ensure consistency between transit pathbuilding parameters and mode choice coefficients, among other things.
The number of stops on the tours should be consistent with the modes used. For example, walk
and bike trips tend to have fewer stops, probably due to the increased travel time required to
access activity locations. Tours in which there are shared-ride trips often have more stops than
other modes, due to pick-up/drop-off of passengers. Tours with drive-transit modes also typically
have more stops in fact, one of the motivations for driving to transit is so that the car is
available to perform out-of-home activities before or after work.
824
The location of intermediate stops on the tour should be reasonably related to the location of the
tour origin and primary destination. Research indicates that travelers generally seek to minimize
overall travel time by chaining trips in a logical order. For example, if a traveler needs to stop for
groceries on the way home for work, a store is typically chosen near the home or en route
between work and home, as opposed to somewhere in the opposite direction from home.
Similarly, the locations of stops on tours are influenced by the mode used for trips on the tour. If
the traveler stops on the way to work by transit, the stop location will likely be within walking
distance of home or work, or in some cases at a transfer location between home and work. It is
unlikely that the traveler will choose a location that is not served by transit or within walking
distance of their origin or destination.
So, how can we ensure consistency across these decisions on at a tour-level?
One aspect of consistency relates to the structure of the model chain, and what choices are
modeled explicitly versus determined implicitly. For example, model system with an explicit
ride-sharing model for escort tours will explicitly predict pickup/drop-off stops on tours, thereby
ensuring that stop locations and trip modes are consistently determined across persons in the
household, with respect to location (with stops modeled at schools) timing (with all departure
and arrival times consistently modeled across tour participants), and mode (with shared-ride
modes for appropriate trips on tours).
Constraints and situational variables are also used to ensure consistency. An example of
constraints is the use of tour mode to constrain the modes used for trips on a tour. For example, if
a traveler chooses walk-transit to work, that traveler cannot use their personal vehicle for
intermediate stops on the tour. Therefore drive-alone is prohibited at the trip level for walktransit tours. Situational variables (variables that are the results of previously applied models) are
also used; for example, auto ownership levels influence mode choice. Travelers in zero auto
households are more likely to walk, bike, and take transit than multiple-vehicle households.
Finally, log-sums are used in choice models to ensure consistency. For example, trip mode
choice log-sums are used as explanatory variables in intermediate stop location choice, to ensure
that chosen stop locations are consistent with available trip modes.
825
Page 8
Key Concepts
Work Tour
Zone 1
Work-Based Tour
Zone 3
Zone 4
Zone 2
This Page shows a simple example of how a situational variable can be used to ensure
consistency. In this case, the situational variable is the chosen mode to work, which is walktransit. This variable is used to constrain the alternatives available (and the probabilities of the
available alternatives) in the work-based sub-tour mode choice model. If a traveler arrives at
work by transit, they are unlikely to drive-alone to lunch. In most cases, we can eliminate drivealone as a potential tour mode for lunch. Additionally, shared-ride is an unlikely choice for this
tour, which is reflected by highly negative alternative-specific constants for shared-ride. A much
more likely choice is walking or walk-transit to lunch. This is one of the reasons why downtown
workers tend to walk to lunch they do not have their car available for their lunch trip.
Additionally, the choice of a restaurant for lunch should be affected by the mode to work; if an
auto is not available, the traveler should choose a lunch-place relatively close to work (within
walking distance). This is ensured by the use of a destination choice log-sum in the at-work subtour which reflects the chosen mode to work. If the chosen mode to work is walk-transit, the logsum will reflect a much steeper decay with respect to distance.
826
827
Page 9
Auto ownership
Transit pass
Trips
Destinations
Modes
Synthetic
Population
Mobility
Choices
Model
Outputs
Activity-Based Modeling: Tour Mode, Primary Destination,
Intermediate Stop Location and Trip Mode
Model
Inputs
Long-Term
Choices
Daily Activity
Pattern and
Tours
Trip
Assignment
Work location
School location
Tours
Purposes
Schedule
Highway volumes
Transit boardings
While ABMs can vary in structure, this diagram shows the location of tour & trip detail choices
(tour mode, primary destination, intermediate stop location and trip mode) in a typical model
stream. The text on the outside shows the types of outcomes predicted by each model stage.
When we are ready to predict tour and trip details, we already have a synthetic population of
households and persons with their home locations, we have predicted the primary work and
school locations, auto ownership and other mobility decisions, and we have generated and
scheduled tours using a daily activity pattern model. We do not yet know the primary destination
of any non-work and non-school tours, the tour mode, the location of intermediate stops, or the
trip mode. Once we are able to fill in these details, we are ready to convert the simulation data
into trip tables that can be assigned to the network.
828
Page 10
Home
Similar to
mandatory tour
destination choice
Probability
depends on a size
term, and an
impedance to each
TAZ
Singly-constrained
or doublyconstrained?
10
Remember, that we have already determined the primary destination of work and school tours as
part of the long-term choices. The method for predicting the primary destination of nonmandatory tours is the sameto use a destination choice model.
In this example, we consider a traveler picking a location to go shopping. They consider the
number of retail opportunities in each TAZ (the size), and the impedance to get to each TAZ.
TAZs that have many shopping opportunities and are easy to get to will have the highest
probability of being chosen. The impedance considers the cost of going there AND back.
Would opening a store in a highly accessible location or an inaccessible location generate more
traffic? Depending on how one answers this question, one can choose to use either a singlyconstrained or doubly constrained destination choice model. A singly-constrained model refers to
one in which the only constraint is on tour origins, while a doubly-constrained model refers to
one in which the total tour destinations in each zone are equal to the input tour attractions in that
zone. In order to reflect the influence of accessibility on total tour attractions, we often allow
829
830
Page 11
Columbus
Atlanta
Bay Area
Sacramento
Generation
Generation
Generation
Generation
TOD
Destination
Destination
Destination
Destination
Mode
TOD
Mode
Mode
TOD
Mode
TOD
SF
11
The sequencing of time-of-day choice relative to destination and mode choice is different in
various model designs. It is ok that we dont know the mode when were picking a destination,
because we can use a mode choice log-sum, which measures the composite impedance across all
modes.
What do we do if we dont know the time periods when picking a destination?
Assert fixed time periods
Iterate between time-of-day and destination choice
Use logit averaging across multiple periods (a time-of-day log-sum)
Many of these issues have already been addressed in the webinar on accessibilities (Webinar 6),
and you can refer to that webinar if you need a refresher.
831
Page 12
Model Structures I
Individual
Daily Activity Pattern
(Tours and presence
of stops by type
Tour Primary Destination
Tour Primary Mode
DaySim Structure
Daily activity pattern models
predict 0 vs. 1+ stops on
outbound and inbound portions
of tours
Tours with intermediate stops are
more likely to choose auto modes
Stop Frequency
(Exact Number)
Stop Location
Trip Mode
Activity-Based Modeling: Tour Mode, Primary Destination,
Intermediate Stop Location and Trip Mode
12
This diagram provides a little more detail on the DaySim model structure currently implemented
in Sacramento. The model first generates individual daily activity patterns, defined by a primary
tour purpose and by whether the tour will have one or more stops on its outbound half, its
inbound half, or both halves. This is followed by choice of locations for the primary activity
location (if a non-mandatory purpose), and tour mode.
One advantage to this approach is that, if it is known that there will be some intermediate stops
on a tour, an individual is more likely to choose the auto mode. However, DaySim then models
the exact number and purpose of stops at the trip level after all tour-level choices for the tour.
The model is a stop-repeat model, in which repeat consists of an alternative for each purpose.
Once the exact number of stops by purpose has been determined for each half tour, locations and
modes are chosen for those trips.
832
Page 13
Model Structures II
Coordinated
Daily Activity Pattern
Explicit
Ride Sharing
Tour Primary Mode
Stop Frequency
(Exact Number)
Stop Location
CT-RAMP Structure
Tour mode choice occurs before
stop frequency
Auto tours are more likely to have
intermediate stops than walktransit tours (drive-transit tours
tend to have as many, if not
more, stops as auto tours)
Explicit ride-sharing for escorting
kids to school
Trip Mode
Activity-Based Modeling: Tour Mode, Primary Destination,
Intermediate Stop Location and Trip Mode
13
This diagram shows a different model structure, representing the CT-RAMP design. Here,
explicit ride-sharing arrangements are modeled prior to choosing a tour mode, so it is known in
advance whether a car will be needed and if it will contain multiple occupants, in this case
household members. This is the main difference between this model and the version of DaySim
just shown. The actual number of stops on the tour is not determined until after tour mode
choice. Both model systems assume that the tour mode is known first and conditions the actual
number of stops on the tour. Similar to DaySim, this is then followed by stop location and trip
mode choice model application.
833
Page 14
Shared-Ride
Zone 4
Drive-Alone
Zone 3
14
Now lets look at how tour mode can be defined. In some cases, all trips on a tour have the same
mode. For example, if a traveler drives alone from home to work and back, there are two drivealone trips. It is also common to observe mode switching on a tour, and for that reason it cannot
be ignored. For example, we often observe parents escorting children to school and other events,
on their way to/from work or other activities. This is typically the most common type of tour in
which we observe mode switching.
This Page shows an example of such a tour. A worker leaves home with a child, drops that child
off at school (in zone 4) and then continues on to work alone. At the end of the day, the worker
returns to school to pick up the child and drives home. In this case we observe one school tour
with no mode switching (the child school tour with two shared-ride trips) and one work tour with
an intermediate stop on the way to work and one intermediate stop on the way back home. The
first and last trip of the workers tour is shared-ride, and the second and third trips on the tour are
drive-alone. If mandatory escort tours are modeled explicitly, the modes of the child and worker
tours are determined explicitly based upon the persons participating in the tour. But when they
are not, we would typically assign a tour mode based upon the highest occupancy mode
occurring on the tour. In this case, the highest occupancy mode on the tour is shared-ride, so the
834
tour mode is shared-ride. The reason for doing so is so that we can let the presence of highoccupancy vehicle lanes influence the location of intermediate stops. However, we allow drivealone to occur on the tour as well. We shall see later how this is addressed in trip mode choice, or
trip switching models.
835
Page 15
Zone 3
Trip 3: Walk
Zone 2
15
Here we see a typical transit tour, where the traveler takes bus to work, and returns home by bus
with a stop for groceries. In this case, the intermediate stop is followed by a walking trip home.
We would define the tour mode for this tour as walk-transit, with two walk-bus trips and one
walk trip. In this case, wed want to find a shopping stop location that was close to a bus stop, on
the route between work and home.
836
Page 16
Zone 1
Trips 1 and 4:
Shared-Ride
Trips 2 and 3:
Local bus
Zone 3
Which is it?
Transit tour with stop (auto): Tour mode = Drive-transit (P&R)
Transit tour with stop (auto): Tour mode = Drive-transit (K&R) for one
person; drive (auto) for the other.
Activity-Based Modeling: Tour Mode, Primary Destination,
Intermediate Stop Location and Trip Mode
16
This tour type is a little more complicated because it involves both shared rides and transit. As
drawn, this could be either a park-and-ride or kiss-and-ride situation. If multiple persons share a
ride to a park-and-ride lot and then take the bus to their respective jobs, and then come home
together, then this is a park-and-ride situation for both travelers. On the other hand, if one person
is being dropped off (kiss-and-ride) and then goes to work on the bus, then it is a kiss-and-ride
(drive transit) trip for the worker, but the person who dropped them off may not be using transit
at all, and may have a different pattern.
For this type of complicate pattern, one needs to look at the full tour patterns for both persons
and determine whether they are both taking bus after arriving in Zone 4. In addition, the return
trip home could be either the same or different.
837
838
Page 17
SR2
SR3+
Tour Mode
Walk Bike
Drive-alone Free
Y
Y
Y
Drive-Alone Pay
Y
Y
Y
Shared-Ride 2 Free (GP Lane)
Y
Y
Shared-Ride 2 Free (HOV Lane)
Y
Y
Shared-Ride 2 Pay
Y
Y
Shared-Ride 3+ Free (GP Lane)
Y
Shared-Ride 3+ Free (HOV Lane)
Y
Shared-Ride 3+ Pay
Y
Walk
Y
Bike
Highest occupancy
Walk-Local Bus
trip mode identifies
Walk-Express Bus
auto tour mode
Walk-Bus Rapid Transit
Walk-Light Rail Transit
Walk-Commuter Rail
Activity-Based Modeling: Tour Mode, Primary Destination,
Intermediate Stop Location and Trip Mode
WalkTran
Y
Y
Y
Y
Y
Y
Y
PNRTran
Y
Y
Y
Y
Y
Y
Y
Y
Y
KNRTran
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Presence of
walk-transit
identifies walktransit tour
17
Now that weve seen some examples of tour and trip mode definitions, lets look at a taxonomy
for defining tour modes and trip modes. This correspondence table is taken from the SANDAG
activity-based model. It shows the allowable trip modes (rows) for each tour mode (columns).
You can see that the trip modes have a lot more detail than the tour modes, and that multiple trip
modes are allowed for each tour mode.
As previously shown, for auto tours, the highest occupancy trip mode identifies the tour mode.
Tours with walk-transit trips (and no drive-transit trips) are identified as walk-transit tours (even
though there may be shared-ride trip legs on these tours). Tours with PNR-transit trips are
identified as PNR-transit tours, even though there may be both drive-alone and shared-ride trip
legs on these tours. KNR-transit tours are similarly defined (though drive-alone is typically
disallowed on these tours).
839
840
Page 18
HOV\Managed Lanes
Light-Rail
Heavy Rail
Commuter Rail
18
The activity-based mode choice model can consider any number of modes, just like a trip-based
mode choice model. These range from more simple definitions (like auto, transit) to more precise
descriptions such as managed lanes (including high-occupancy vehicle, high-occupancy toll, and
toll lanes), light-rail transit, bus rapid transit, heavy rail and commuter rail.
High-Occupancy Vehicle (HOV) Lanes are lanes in which vehicles are restricted based on
occupancy; typically 2 or more persons are required, although some lanes require 3 or more
persons. High-occupancy toll lanes are lanes in which vehicles are tolled based upon occupancy.
For example, single-occupant vehicles pay a toll while multiple occupant vehicles go free. Busrapid transit is a type of transit mode featuring buses that offer operating characteristics similar
to rail; these include separate rights-of-way, rail-like station amenities such as covered platforms
and/or rider information, and low-board vehicles for easy boarding and alighting. Light-rail
transit is a type of rail transit that can operate in either separate right-of-way or in mixed-flow
with auto traffic. Heavy rail consists of passenger rail cars operating singly or in trains of two or
more cars, on rights-of-way from which all other vehicular and foot traffic is excluded.
841
Commuter rail is passenger train service operating between a central city, its suburbs, and/or
another central city.
842
Page 19
Nonmotorized
Transit
Drive alone
Walk
Walk
access
Shared ride
2
Bike
Drive
access
Shared ride
3+
Activity-Based Modeling: Tour Mode, Primary Destination,
Intermediate Stop Location and Trip Mode
19
This Page shows the San Francisco tour mode choice model structure. As you can see, the tour
modes are fairly aggregate at the tour level. The auto mode is broken out by occupancy, and
transit is broken out by mode of access (walk versus drive, which includes both park-and-ride
and kiss-and-ride). There are also walk and bike modes.
843
Page 20
Shared Ride
Shared Ride
2 (S2)
NonMotorized
Transit
Shared Ride
3+ (S3)
Drive to
Transit (DT)
Walk to
Transit (WT)
Full Highway
Network
Full Highway
Network
Local Bus
Local Bus
Non-toll
Highway
Non-toll
Highway
Light Rail
Light Rail
Premium
Bus
Premium
Bus
Tour Mode
Walk (WK)
School Bus
Bike (BI)
Path Type
20
This Page shows the tour mode choice structure for the SACOG activity-based model,
SACSIM. This structure shows a bit more detail and includes path type choices below the tour
model level. For auto modes, this includes a choice between the full highway network and the
non-toll highway network. For transit modes, this includes choices between local bus, light rail,
and premium bus.
844
Page 21
Nonmotorized
Auto
Drive alone
Shared ride
2
Shared ride
3+
GP
GP
GP
Pay
Pay
Pay
Walk
Bike
Transit
Walk
access
Drive
access
Local bus
Local bus
Light rail /
ferry
Light rail /
ferry
Express
bus
Express
bus
Path Type
Heavy rail
Heavy rail
Commuter
rail
Commuter
rail
Tour Mode
21
This Page shows the tour mode choice model for the Metropolitan Transportation Commission.
Here there are even more transit modes differentiated.
845
Page 22
Nonmotorized
Auto
logsums
Pay
GP
HOV
Pay
GP
Walk
Bike
Walk
access
PNR
access
School Bus
KNR
access
Local bus
Local bus
Local bus
HOV
Express
bus
Express
bus
Express
bus
Pay
BRT
BRT
BRT
Tour Mode
Path Type
Mode Choice Nesting
for Logsum
Activity-Based Modeling: Tour Mode, Primary Destination,
Intermediate Stop Location and Trip Mode
LRT
Commuter
rail
LRT
Commuter
rail
logsums
GP
Shared ride
3+
logsums
Shared ride
2
logsums
Drive alone
Transit
LRT
Commuter
rail
22
And here is the mode choice model for San Diego. Note that the path type choice (for example,
local bus versus express bus, etc.) is considered by the model by taking a log-sum across all
transit path choices, while the tour mode is defined by the more aggregate walk, PNR, or KNR
access mode. In other words, trip mode choice is only conditional on the choices allowed for a
walk-transit tour, instead of the more limited choices that would be allowed on a walk-local bus
tour. This way, travelers have freedom to switch between transit line-haul modes between
multiple trips on their tour.
846
Page 23
23
Here are typical inputs to tour mode choice models. Note that round-trip in-vehicle time, wait
time, access & egress time, and costs for each mode are taken into account. This allows the mode
choice models to be sensitive to round-trip levels of service. For example, an improvement of
transit in the evening will increase transit ridership in the morning period as well, to the extent
that some tours that are ending in the evening period actually start in the morning (and evening
service cutbacks will have the inverse effect on morning ridership). The models are also sensitive
to income, gender, age, and other household and person level variables. Land-use and urban form
variables such as intersection density, land-use mix, presence of sidewalks, etc., can also be used
in tour mode choice models, particularly affecting the probability of walk trips and walk-transit
trips. Situational variables, which are predicted by models higher in the model chain, also have
an effect on tour mode such as auto ownership, transit pass ownership, toll transponder
ownership, and free parking eligibility\parking cost models. Tour purpose, joint travel, number
of stops, and other tour variables can also be used in tour mode choice.
847
848
Page 24
Time
Cost
Distance
Distance traveled, cost on Managed Lane
Reliability (e.g. difference between free-flow and congested time)
24
A key input to mode choice models are travel time and cost skims, or level of service matrices,
for every zone-pair and time period in the model. Some models also segment skims by income or
other socio-economic market, as described below. Most practical activity-based models rely
upon static equilibrium assignment methods to produce travel skims.
Auto skims typically include time, cost and distance. One set of skims is prepared for each auto
occupancy level, to reflect differences in access to HOV lanes and toll costs. Additionally, if the
model distinguishes between free and toll paths, skims must be prepared for free paths where toll
lanes in the network are turned off, and another set of skims is prepared for toll paths where all
pay facilities are turned on. For such models, it is often useful to also skim the distance traveled
on managed or pay lanes for each OD pair, as well as the toll cost. It has been shown that
transport system reliability is a key factor on traveler decision-making, and therefore it is
increasingly common to use some measurement of reliability in mode choice. However,
reliability is difficult to measure with a static equilibrium model. One example measurement of
reliability uses the difference between free-flow and congested travel time as an indication of the
849
variability of travel time. Income is often used to segment skims where toll lanes play a role, in
order to take into account value-of-time on toll path choice (more on this later).
Transit skims include in-vehicle time, typically segmented by line-haul mode so that the model
can distinguish between different services available to the traveler. First and transfer wait times,
along with number of transfers, describe the frequency of transit service and whether transfers
are required. Access, egress, and transfer walk and drive times describe proximity of transit to
the trip origin and destination. Finally, transit fare is skimmed. The quality of access and egress
times is highly dependent on the level of spatial aggregation in the zone system, where trip ends
in large zones have the most error in terms of access and egress times. Some models, such as
Sacramento and Denver, compensate for this error by replacing the zonal level estimate of walk
access and egress time with the time between the origin and destination parcel and the closest
transit stop. The San Diego model uses a novel approach to measuring transit times which we
will see next.
Skims must also be provided for walk and bike modes. Often distance on different types of
facilities are skimmed and help measure the quality of bicycle trips.
850
Page 25
Transit Path-Building
Different Origin MGRA
(same TAZ) has different
walk & transit times
Boarding TAP
requires bus transfer
to rail
One of the problems in skim-building is that travel skims consume memory and disk space,
because a set of matrices is built for every zone pair and time period. The amount of memory and
space required to store skims increases squared times; for example, the space required for a
5,000 zone system is 4 times greater than that required for a 2,500 zone system. When the spatial
system is highly disaggregate, such as when parcels or micro-zones are used, it is simply not
possible to skim the network for each geographic pair. Various methods are used to get around
this problem. For example, transportation analysis zones are used for auto skims, since the
relative error related to larger TAZs is fairly small at auto speeds. Walk and bike modes,
however, suffer from greater error with respect to spatial aggregation, since these modes are
slower. The Sacramento and Denver models use parcels for the representation of space, and
represent walking and biking times at the parcel level for close-in parcels. Transit access and
egress is skimmed at the zonal level, but the nearest bus stop is assumed consistent with skims
and time from the parcel to the bus stop is used to replace the skimmed walk time.
851
The San Diego model employs a novel approach to transit path building in which transit stops
are represented explicitly in the transit network. Transit skims are built between stops instead of
between zones, and micro-zones, or Master Geographic Reference Areas (MGRAs) are used to
represent the origins and destinations in the model. Walk time is measured between the microzone centroid and the actual stop, provide an accurate description of the level-of-service between
various stop pairs between the origin and destination. This Page shows an example of the
approach. The origin MGRA has three stops that provide access to the destination. The stop to
the north provides access to a bus line, with no transfers required. The stop immediately to the
south provides short walk access to a feeder bus which incurs a transfer at rail. Alternatively, the
traveler can walk further to rail and forego the local bus transfer. A different MGRA in the same
zone has different access and egress options, where the direct rail access is more attractive due to
the shorter walk. The San Diego model builds a utility for each path and these path utilities are
used in the mode choice model to influence both mode and path choice.
852
Page 26
Land Use
Pattern-specific
Demographics
Interaction
26
Here is a list of the explanatory variables specified in the tour mode choice model of the
Sacramento activity-based model. At the top are the usual level-of-service variables, including
in-vehicle and out-of-vehicle travel times, and costs for auto operations, tolls, and transit fares.
Below that are land use data variablesmixed use density and intersection density at the
destination.
This is followed by a couple of day-pattern pattern-specific variables. First, there is an escort
stop dummy variable which, in the absence of explicit household rides, increases the propensity
to choose shared ride modes. In addition, there is a variable indicating the number of stops on the
tour (other than escort), which increases the probability of choosing auto modes.
Demographic variables also play a major role in endogenous segmentation of tour mode choice.
Here, we have listed household auto sufficiency based on cars to workers, income groups,
household type defined by single persons and presence of children. In addition, being an
individual-based model, we have information on person attributes, such as age and gender.
853
Finally, it is common to have interaction terms. For example, the work-based sub-tour mode
choice model considers the tour mode chosen for the primary work tour. This strongly conditions
the mode to be used on the sub-tour and places constraints on what is available. In addition, there
are dummy variables that indicate biasesfor example, we do not have a completely separate
shopping mode choice model, because shopping is grouped within the home-based other (nonmandatory tour mode choice model. However, we include a shopping indicator variable in that
model that indicates that shoppers are more likely to choose shared ride than persons whose
primary tour purpose was something else.
854
Page 27
27
27
In addition to traditional variables like travel cost, purpose, time-of-day, and household
characteristics, activity-based models allow us to use a richer set of variables, and have extended
our knowledge and understanding of travel behavior. Here are some new variables available in
the AB modeling framework:
Page 28
28
Typically household survey data is used for mode choice model estimation and calibration.
However, often household surveys have too few transit observations upon which to estimate
and/or calibrate transit choices. Therefore on-board surveys are used in conjunction with
household survey data, just as is done to estimate trip-based mode choice models. One issue with
the use of on-board surveys is that the surveys are typically origin-destination based. That is,
they only ask about the trip that the traveler was observed taking when intercepted on the transit
vehicle. In order to be useful for activity-based model development, additional questions should
be asked that attempt to identify the ultimate tour purpose, tour origin\destination locations,
and/or tour mode. Some example questions are given on the next Page.
856
857
Page 29
51%
13%
10%
12%
12%
2%
100%
29
Here are some questions that can be added to an on-board survey to better understand the tour
context for the reported origin-destination trip. Home and work (and school) address are useful
for identifying the tour origin and primary destination locations. Questions can be added to better
identify tour purpose, such as whether the traveler has been or is planning to go to work (or
school) on this tour. Questions can also be added to better identify tour mode, such as how the
traveler plans to get back home.
The table on the Page shows the cross-tabulation of four-step model trip purpose versus tour
purpose. In particular, it shows how 25% of non-home-based trips are made on work tours. A
significant portion of Home-based Other and Home-Based Shop trips are also made on work
tours.
858
859
Page 30
30
The parameters of tour and trip mode choice models are typically estimated using maximum
likelihood estimators available in logit model estimation software. In maximum likelihood
estimation chosen alternatives are compared to non-chosen but available alternatives, and a set of
parameters are found which maximizes the probability of selecting the chosen alternatives.
It should be noted that parameter values should be consistent between tour and trip mode choice.
That is, tour time and cost parameters should generally be smaller (approximately half the size)
of trip mode choice model time and cost parameters. This is because the tour models take into
account round-trip levels of service while the trip mode choice models take into account only the
trip level of service. In order to ensure consistent elasticities, the parameters in the tour model are
approximately half the size of the trip model parameter (as shown in the equations at the end of
the Page). Similarly, alternative specific constants at the tour level are approximately twice as
high than constants at the trip level (expressed in equivalent minutes), since they represent nonincluded attributes of alternatives which is being taken across at least two trips.
860
861
Page 31
Probability Density
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
$-
$5
$10
$15
$20
$25
$30
31
A key advantage of activity-based models is the use of individual parameter variation to reflect
unobserved heterogeneity in the sensitivity to time and or cost. Such heterogeneity is very
important when modeling pricing alternatives, to eliminate the aggregation bias associated with
the use of average values-of-time. This Page shows values of time that vary probabilistically
within household income levels. These distributions were estimated from a combined stated and
revealed preference survey conducted in San Francisco. The data were collected as part of a
pricing study and was used to enhance the San Francisco activity-based model system to address
road pricing alternatives. Each simulated person in the San Francisco activity-based model
selects a value-of-time randomly from the distribution corresponding to their household income.
This value-of-time is converted into a travel cost parameter that is used for all travel models
including mode choice. The curves reflect the typical value-of-time distributions observed in
data --for each income group, there are some travelers who have a much higher willingness-topay than the average for their income group. Higher than average willingness to pay results from
schedule constraints, personal preferences, and other unobserved attributes.
862
863
Page 32
Calibration methodology
Things to check:
Tour Mode
Drive-Alone
Shared 2
Shared 3+
Walk
Bike
Walk-Transit
PNR-Transit
KNR-Transit
School Bus
Total
Observed
Auto Sufficiency
No
Vehicles< Vehicles>=
Vehicles
Adults
Adults
138,616
544,877
7,307
58,993
125,455
4,201
30,976
91,925
6,058
12,612
8,102
2,636
3,632
4,072
11,995
9,847
10,368
223
1,278
3,621
211
1,083
1,211
32,632
257,038
789,632
Total
683,493
191,755
127,102
26,773
10,340
32,210
5,123
2,506
1,079,302
Estimated
Auto Sufficiency
No
Vehicles< Vehicles>=
Vehicles
Adults
Adults
0
138,585
552,000
7,520
58,990
127,260
4,330
31,035
93,300
6,305
12,615
8,375
2,715
3,665
4,185
12,475
9,945
10,685
0
1,315
3,830
220
1,120
1,275
33,565
257,270
800,910
Total
690,585
193,770
128,665
27,295
10,565
33,105
5,145
2,615
1,091,745
32
Just like trip mode choice models, tour mode choice models are calibrated to reproduce
aggregate shares obtained from observed data. Often, tour purpose and auto sufficiency are used
to segment alternative-specific constants. Rather simplistically, calibration means adjusting the
value of these constants until the model forecast matches the observed shares. In reality, proper
calibration often requires revising upper level models and skim building procedures, in addition
to revising the value of the model constants. Here we see observed versus estimated work tours
by tour mode and auto sufficiency for San Diego.
864
Page 33
Work-Based Sub-tour
33
This Page shows tour mode choice model calibration results for the Sacramento activity-based
model by tour purpose. As you can see, it is quite easy to match calibration target values for
mode shares in the aggregate.
865
Page 34
34
It is also important to check the calibration of the model at more disaggregate levels. This Page
shows scattergrams comparing estimated to observed walk-transit and PNR-transit tours by tour
origin and primary destination district. The walk-transit comparison looks close to observed,
while the PNR-access scattergram reveals some differences which must be further investigated.
866
Page 35
Directional Pricing
EV
AM
Considers
round-trip
Working late. Can I take the bus to work?
Activity-Based Modeling: Tour Mode, Primary Destination,
Intermediate Stop Location and Trip Mode
Cost applied to
Tour Decisions
Pricing Boundary
Here are some examples that illustrate the importance of accounting for tour-level conditions on
the choice of mode for each trip. The types of effects that are shown here are impossible to
capture with a trip-based model.
In the first example, an improvement in evening transit service is reflected in the choice of mode
for morning trips. As expected, improving evening transit service, for example by expanding the
span of express bus service, results in increased evening transit patronage. It also results in
increased morning patronage because workers with long hours at the office can take transit to
work and rely upon better transit service to return home. That is, improving the level of service
for just one leg of the tour affects the choice of mode for the entire tour.
In the second example, introduction of pricing on the way to work influences not only the mode
to work, but potentially the location and mode of intermediate stops between work and home.
For example, persons avoiding tolls will follow a different path to work that will offer a different
level of access to discretionary activity opportunities and locations. By the same token, persons
who choose a tolled path will likely remain on that path for a longer period of time (assuming
867
there is a time-savings incentive) and potentially bypass activity opportunities and locations that
would cause them to deviate from that path.
868
Page 36
869
Page 37
CT-RAMP Structure
Individual
Daily Activity Pattern
(Tours and presence
of stops by type
Coordinated
Daily Activity Pattern
Explicit
Ride Sharing
Tour Primary Mode
Stop Frequency
(Exact Number)
Stop Location
Trip Mode
37
Both DaySim and CT-RAMP choose stop location after tour mode and before trip mode. What is
known at this point in the model system?
What is unknown at this point in the model system is the location of each stop on the tour, which
will be predicted by the intermediate stop location choice model, and the mode used for each trip
on the tour, which will be predicted by the trip mode choice model.
870
871
Page 38
Quality variables
38
Before continuing with intermediate stop location choice models, lets revisit how destination
choice models work. A destination choice model takes into account spatial separation between
activity locations, as well as the amount of opportunities to engage in the activity in each
destination. The number of opportunities in the zone is referred to as the size term in the model.
The size of the zone is logged, so that the probability of selecting an alternative is equal to the
relative size of the alternative compared to all other alternatives, all else being equal. The spatial
separation between the origin and the potential destination is represented by time, distance,
and/or mode choice log-sum terms. The process used for modeling spatial separation for
intermediate stops is a bit different than that used for tour primary destination choice.
872
Page 39
eVnc e
Vnj
eVnk
Zone c
Zone j
Zone k
39
Why do need to create a non-linear size term? Why not just include these quantity variables in
the regular utility expression? The size term helps us to avoid one aspect of the modifiable areal
unit problem, where arbitrarily drawing boundaries around an area unit (like a TAZ) can affect
statistical analysis of that unit. If we split a zone into two (or more) smaller zones, we expect the
summed probability of choosing the two split zones to be equal to that of the original single zone
(all else being equal). For a logit model, this holds when the exponentiated utilities of the two
split zones sum to the same value as that of the exponentiated utility of the original combined
zone.
873
Page 40
Thus, we have
40
If we specify representative utility as a natural logarithmic function, then the exponentiation will
reverse the log function, and we can insure this equality. If we then let the impedance function
(distance, travel time, generalized cost) be a negative exponential function, then we have a utility
expression that is a combination of a logged attraction term and a linear-in-parameters
impedance term. We refer to this attraction term as the size of the zone, because it represents a
positive quantity. Since there may be many reasons to be attracted to a zone, for example both
dining and shopping, we may have multiple attraction variables within the size function.
As a practical note, I should mention that the beta term is typically not estimated and fixed to 1.0
in order to avoid complications with interpretation. In addition, at least one of the size variables
must also have its parameter fixed to 1.0 (theta) due to parameter identification restrictions in
estimation. A third restriction is that the estimated thetasthe size variable coefficients are
restricted to have positive values in keeping with the theory that the variables represent positive
quantities in which more is considered better (more attractive).
874
875
Page 41
Primary
Costij
Origin i
Costik
Destination j
Costkj
Intermediate
Stop k
41
876
Page 42
Primary
Origin i
Destination j
Costij
Costkj
Intermediate
Stop 1
Costik
Intermediate
Stop k
42
This Page shows the same calculation for a tour with two stops. In this case, the first
intermediate stop location has been determined, and we are now choosing the second
intermediate stop location. The out-of-direction travel cost in this case is based upon the
previously-chosen stop location as the origin and the previously chosen tour primary destination
as the destination.
877
Page 43
Primary
Origin i
Destination j
Costij
Costkj
Intermediate
Stop k
Costik
Intermediate
Stop 1
43
This Page looks very similar to the previous one, with one important difference. The previous
Page reflected the approach used in CT-RAMP.
An important difference between DaySim and CTRAMP is that DaySim models tour destination
arrival and departure times at the tour level, whereas CTRAMP models origin departure and
return times at the tour level. This is a somewhat fundamental difference that has significant
implications for the trip-level models. While the method of calculating cost is analogous, in
DaySim the first stop added is "adjacent" to the tour destination. When a second stop is added, it
is added backwards between the first intermediate stop and the origin. The implication is that
arriving on time at the primary destination and leaving on time and is more important, and that
additional stops will serve to push back/forward home departure times and arrival times.
It is unclear whether this offers any behavioral or practical advantage. For example, one could
argue that home departure and arrival times are more important under many circumstances.
878
Page 44
Other
44
Other time and cost terms can be considered in intermediate stop location choice models. For
example, it has been found that the distance between the tour origin and the tour primary
destination influence the stop location, particularly for long tours, since people tend to be more
familiar with areas immediately around their home and workplace and are therefore more likely
to choose stop locations near these habitual locations. Trip mode choice log-sums (or generalized
costs consistent with the tour mode) are often used instead of distance or time, so that the stop
location reflects accessibility according to the chosen tour mode. For example, if the chosen tour
mode is walk-transit, one can use a log-sum taken across trip mode choice (which will reflect a
higher waiting for transit and non-motorized modes) or use transit time, walk time, or some
average of the two to represent accessibility.
Other explanatory variables can include household and person demographics, land-use and/or
urban form, the purpose of the tour, and alternative-specific constants such as river crossings
though these constants should be used with caution.
879
880
Page 45
Escort
Shop
1.00
Tour Purpose
Maint
Eat
1.00
0.85
1.00
Visit
0.33
Discr
0.22
0.02
0.03
0.14
2.46
1.00
0.72
1.00
0.55
1.00
0.65
0.44
45
Size terms are equivalent to trip attraction equations in a gravity model. Here we see the size
terms for stop purposes used in the San Diego activity-based model. Each stop purpose is shown
as a column, and micro-zone variables are shown as rows. Each cell has the coefficient for the
size term variable. For each tour purpose one of the variables is arbitrarily chosen as the base
variable, and its size term coefficient is set to1.0. This allows one to compare the effect of other
variables relative to the base variable. For example, lets look at the escort purpose; each school
enrollee is equivalent to 44% of a household, in terms of attractiveness of stop locations.
881
Page 46
Sampling approaches
Nave choose n alternatives at random
Intelligent - based on simplified model
Availability constraints
Size term > 0
Available according to tour mode
For walk-transit tours, must be able to get there via
transit or walking
For walking tours, stop must be within walking
distance of both tour origin and destination
46
Often sampling is required for intermediate stop location choice estimation and application, due
to the number of alternatives considered. This is especially true if the model system utilizes a
highly disaggregate spatial system, such as micro-zones or parcels, where the number of
alternatives ranges in the tens of thousands. In such cases, it would be computationally infeasible
to calculate probabilities for each potential destination. Alternatives sampling first requires
creating a selection set, then computing a utility for each sampled alternative and finally
selecting an alternative from the sample.
There are number of alternative approaches to sampling, ranging from nave to intelligent. A
nave approach would be to simply selecting n alternatives at random, and relying upon the
number of alternatives to ensure that at least a subset of alternatives is reasonable for the tour. A
more intelligent approach involves the use of a simplified destination choice model to generate a
probability of inclusion for every potential destination, selecting a subset of alternatives
according to that probability distribution, and then applying the full model to the sample set. The
882
intelligent approach provides a more realistic choice set, but at the cost of more computational
time required.
Constraints are used to ensure that alternatives in the model are realistic. A simple constraint is
that each alternative must have relevant variables for the specific stop purpose; for example,
zones without retail employment are not available for shopping stops. Another typical constraint
is that stop availability is constrained by tour mode. For example, stops on transit tours must be
available by either transit or walking. For walk tours, the stop must be within maximum walking
distance (3 to 4 miles typically) of both the tour origin and primary destination. These constraints
ensure that intermediate stops can be accessed by the modes allowed and that are reasonable
given the tour mode, the origin and primary destination of the tour, and all other intermediate
stop locations.
883
Page 47
Calibration of
Intermediate Stop Location Models
Stop out-of-direction trip length frequency distribution
Distance from tour anchor trip length frequency distribution
Distance from tour primary destination trip length distribution
Average distances
Work
Shop
35%
35%
30%
30%
Target
25%
Model
25%
20%
TargetAve
20%
Target
Model
TargetAve
ModelAve
ModelAve
15%
15%
10%
10%
5%
5%
0%
0%
0
10
15
Distance
20
25
30
10
15
Distance
20
25
30
47
Intermediate stop location choice models are calibrated similar to gravity models and trip-based
destination choice models. Average trip lengths and trip length frequency distributions are used
to ensure reasonableness of results based upon comparisons to expanded household survey data.
In addition, out-of-direction distance, distance from home, as well as distance to primary
destination can be summarized. Shown on this Page are trip length frequency distributions for
out-of-direction distance for stops on work tours and shop tours. Both cases show a good match
between estimated and observed distributions.
884
Page 48
Variables include
48
Now that weve explored intermediate stop location choice models, lets move on to trip mode
choice, or trip switching models. These models determine the mode for each trip on a tour. They
are referred to as mode switching models because they account for the likelihood of switching
modes on a tour. For example, on a shared-ride tour, it is frequently observed that some trips are
drive-alone and some trips are shared-ride. This model determines which trips are taken by each
mode.
Trip mode choice models are heavily constrained by the chosen tour mode, to ensure consistency
between modes for each trip on the tour. These consistency relationships are defined in the tour
mode/trip mode table shown earlier, and enforced by constraining the availability of trip modes
for each tour mode, as well as alternative specific constants that are segmented by tour mode. For
example, if the tour mode is walk transit, typically drive-alone is not allowed, because the
traveler does not have their car on their tour. This is an example of an enforced constraint on
mode availability. Additionally, shared-ride has a low probability of selection on walk-transit
tours, because of the difficulty of finding a ride and the additional disutility imposed on a driver,
885
but may still be one of the available modes on walk-transit tours. In this case, shared-ride would
have a negative alternative-specific constant for walk-transit tours, to reflect the disutility of the
choice, resulting in a low probability of selection.
Variables in trip mode choice models include typical time and cost variables for each mode, for
the trip origin-destination pair and relevant time period, as described above under tour mode
choice. Land-use and urban form variables, traveler characteristics, and tour mode are all
potential explanatory variables. Also, the sequence of the trip on the tour is an important
explanatory variable, as we shall see next.
886
Page 49
Zone 1
Zone 3
49
The sequence of the trip on the tour is an important consideration for trip mode switching. For
example, shared-ride tours with an intermediate stop for escort involve dropping off other
household members or picking them up on the way to or from some other activity, like work.
When this travel pattern isnt modeled explicitly, the trip mode choice model predicts which trips
are shared-ride and which trips are drive-alone. Typically, the first and last trips on the tour (the
ones to/from the escort activity) are the shared-ride trips, while the inner trips on the tour
(between the escort activity and the individual activity such as work) are the drive-alone trips
as shown on the graphic. Similarly, on park-and-ride tours with intermediate stops, the first and
last trips on the tour are taken via auto (such as working out at the gym after work) while the
inner trips on the tour are transit trips. Walk transit tours with intermediate stops are often
characterized by short walk stops close to home with a longer transit trip to/from the primary
destination, though there is considerable variation in walk-transit tours.
887
888
Page 50
Drive Alone
Shared Ride 2
Shared Ride
3+
Drive to
Transit
Walk to
Transit
Walk
Bike
School Bus
Drive Alone
Drive Alone
Drive Alone
Drive Alone
Walk to
Transit
Walk
Walk
School Bus
Walk
Shared Ride 2
Driver
Shared Ride 2
Driver
Shared Ride 2
Driver
Walk
School Bus
Bike
Walk
Bike
Shared Ride 2
Passenger
Shared Ride 2
Passenger
Shared ride 2
Passenger
Shared Ride 2
Passenger
Shared Ride 2
Passenger
Bike
Walk
Shared Ride 3
Driver
Shared Ride
3+ Driver
Shared Ride
3+ Passenger
Shared Ride 3
Passenger
Shared Ride 2
Passenger
Bike
Shared Ride 3
Passenger
Shared Ride
3+ Passenger
Bike
Walk
Drive to
Transit
Bike
Walk to
Transit
Shared Ride
3+ Passenger
Tour Mode
Walk
50
This diagram portrays the trip mode choice availability structure, where trip mode is conditional
upon tour mode. The logic in this structure is based on observed trip modes for a given tour
mode. It is also somewhat hierarchical, based on what has been excluded in the trip mode choice
set. For example, on a Drive-Alone tour, shared-ride trips have been excluded. Shared-ride trips
for two persons (driver or rider) are available on Shared Ride 2 tours, but Shared Ride 3+ has
been excluded. Shared Ride 2 and 3+ (driver and rider) are available on the Shared Ride 3+
tours. Shared Ride 2 and 3+ as a passenger, not a driver, are also available on drive-to-transit and
walk-to-transit tours, as well as on Walk and School Bus tours.
889
Page 51
Destination choice model predicts parking TAZ for every auto trip to parking
constrained area (CBD)
Explanatory variables
Walk time to destination
Parking cost
Household income
Zone 3
51
Parking location choice is often addressed after stops have been located and trip mode has been
determined, specifically for auto trips to parking constrained areas. In such areas, it is common
for drivers to park in a zone that is not necessarily the same zone as their destination, either
because of parking availability constraints or due to price differences. Travelers trade off the cost
of parking with the time required to walk to their final destination. Parking location choice
models rely upon the availability, spatial distribution, and cost variation of parking supply
around the central business district to improve traffic assignments in downtown areas, and model
policies that constrain parking supply or change parking cost. Parking location choice models are
also useful when modeling the impact of transit circulator projects, particularly on college
campuses where parking is often expensive and highly constrained.
890
Page 52
Calibration methodology
Things to check:
Drive-Alone
Shared 2
1,476,827
254,733
256,731
2,785
-
Generic
Total
Non-Toll/Free
2,385,672
Shared
3+
173,819
78,274
155,109
1,455
-
HOV
8,949
Toll
10,065
41,354
-
1,675
1,331
4,984
59,650
-
PNRTransit
3,309
1,147
1,347
1,389
10,245
-
Local
35,536
Express
4,358
LRT
34,742
Walk
WalkTransit
KNR-Transit
1,650
79
868
3,164
3,759
Commuter Rail
3,571
52
Trip mode choice models are calibrated based upon observed data tabulations from household
interview data and/or on-board transit surveys. Typically one compares trips by tour mode and
trip mode for each purpose, as shown in the Page. Other summaries include trips by mode and
district (though if tour mode choice was sufficiently calibrated, these summaries should be good
without any adjustment), transit trips by number of transfers, and trips by mode and trip length.
Typically, alternative-specific constants for transit line-haul modes should not differentiate
between tour mode or socio-economic market segment.
891
Page 53
Feedback
Primary effects:
Route and mode
choice
Secondary effects:
Time-of-day and
destination choice
Tertiary effects:
Tour\stop
generation and
mobility models
53
Here is a practical, concrete example of how an activity-based model might respond to a network
scenario. In this scenario, we are modeling increased transit service in a congested corridor
during the AM and PM peak periods. Note that this is the same scenario that was described in the
accessibilities webinar. However, in this case we can see a bit more clearly now that we
understand the full model structure.
Obviously, once we code better transit service in our network, the first change one might see is
that level-of-service skims for affected zone-pairs will reflect the improved transit in-vehicle and
out-of-vehicle times. As a result of these network (and skim) improvements, transit has a higher
probability of selection in route choice (transit riders with a choice of routes will choose the
improved transit routes more). Now we know that there are actually two different mode choice
models, both of which take into account the improved transit service. There will be more transit
tours predicted by the tour mode choice models, and more transit trips on those tours as predicted
by the trip mode choice model.
892
The primary destination choice model and intermediate stop models take into account transit
level-of-service via mode choice log-sums, which reflect the transit improvement. These
improvements lead to higher probabilities of selection for zones in the improved transit corridors,
and therefore more primary destinations and intermediate stops in those corridors. The time-ofday choice also takes into account mode choice log-sums that vary by time-of-day. To the extent
that these log-sums reflect increased transit level-of-service in the AM and PM periods, more
travelers will travel in these periods as a result. Note that with feedback, some auto travelers may
shift back into the peak periods since some of them have chosen to switch to transit, freeing up
some capacity. These effects are represented by the feedback loop to the left of the diagram.
The improvement in the destination choice log-sum has tertiary effects on tour and stop
generation models and medium-term mobility models. Households that reside in the corridor
may generate more direct tours with less stops per tour, as they change their travel patterns to
take advantage of the transit service. Households may opt to own fewer cars and more transit
passes as a result of the improvement. All of these potential travel behavior changes are
represented in activity-based models with well-formulated tour and trip mode choice models and
intermediate stop location choice models.
893
Page 54
Primary effects:
Destination
choice
Secondary effect:
Mode choice
54
Here is another example of how an activity-based model might respond to a policy input. In this
case, we are looking at increased land-use density\development in some corridor. This would be
reflected by a change in the land-use input file to the model.
Primary effects of land-use changes are in destination choice. Both the tour primary location
choice model and the intermediate stop model would reflect increased destinations in the
corridor, due to the increase in opportunities. Since work and school (mandatory) location choice
models are constrained to match employment and enrollment respectively, these models would
show a proportional increase in destinations to the increase in jobs and enrollment in the
corridor. Non-mandatory models would also show an increase in destinations; however, since
these models are typically not doubly-constrained, the increase may not be proportional to the
increase in opportunities. The increase in destinations would also be a function of how
convenient the corridor is to households in trip-based model terms, the trip attraction rate for
non-mandatory tours (and especially intermediate stops on tours) is a function of both the
quantity of employment in the zone as well as the accessibility of the zone to households. For
894
intermediate stops, the amount of stop-making will depend on whether the corridor is accessible
to both households and primary destinations or workplaces. A corridor that is en route between a
bedroom community and a major employment center would be expected to show more
intermediate stops than one that was not.
The secondary effects of increased corridor land-use density would be on mode choice; in this
case, the direction of the change would depend on the quality of transit service, whether the
corridor is served by a toll facility, etc. If transit service in the corridor is better than other parts
of the region, then those trips that change destinations would be more likely to choose transit and
therefore overall transit ridership would likely increase. Similarly, if the corridor is served by a
toll facility that offers a relatively good time savings to toll cost ratio, then an increase in toll
trips is likely. Typically, increases in density also increase non-motorized mode share; because
there are more opportunities closer together, trip length decreases and non-motorized modes are
more competitive for shorter trips.
Tertiary effects include time-of-day changes, tour/stop generation changes and mobility model
changes. The magnitude and direction of these changes is highly dependent on the quality of
service of the modes that serve the corridor with increased density and the level of congestion in
the corridor. If the increased density results in a more congested corridor overall, and
improvements in transit service in the corridor are not significant, one might observe changes in
time-of-day of trips in and through the corridor to less congested (off-peak) periods. Similarly,
there may be changes in the quantity of tours generated and the complexity of those tours.
Typically households that are in denser areas generate more direct tours with less stops per tour,
while households that are in less dense areas generate fewer tours with more stops per tour. The
effects on tour and stop generation will depend to a large extent on where households reside with
respect to the corridor of interest. There may also be changes on mobility models including auto
ownership, transit pass holding, etc.
895
Page 55
What are the advantages offered by the activitybased model treatment of tours and stops?
Greater consistency between modes chosen for all trips
on tour
Greater consistency in destinations chosen between
home-based and non-home-based trips
Less aggregation bias in typical variables such as
parking cost, toll cost, access to transit and nonmotorized time and distance
Ability to incorporate additional household, person
land-use and level-of-service throughout the day
Activity-Based Modeling: Tour Mode, Primary Destination,
Intermediate Stop Location and Trip Mode
55
Shown here is a recap of some of the advantages offered by activity-based models, compared to
traditional trip-based techniques.
Activity-based models offer a greater consistency between modes chosen for all trips on a
tour. This is ensured by use of tour mode to constrain trip modes on the tour, use of trip
or path log-sums to influence trip mode, choosing intermediate stop destinations that are
accessible by tour modes, and constraining work sub-tours based upon tour mode.
Activity-based models offer greater consistency in destinations chosen for home-based
and non-home-based trips. Trip-based models operate on each trip purpose
independently, whilst activity-based models constrain intermediate stop locations on the
tour origin and primary destination, ensuring a logical relationship between all three.
Activity-based models have less aggregation bias in commonly-used variables such as
parking cost, toll cost, access to transit, and non-motorized time and distance. The
aggregation bias is overcome in activity-based models through the use of simulationbased modeling of individuals rather than an entire zone of households of a certain types.
896
Activity-based models offer the opportunity to incorporate more variables than trip-based
models, including household, person, land-use variables and level-of-service that varies
throughout the day.
897
Page 56
56
The concept of tour modes conditioning trip modes is well accepted. The concept of rubberbanding to find locations for intermediate stops on tours is also fairly well accepted. Much of
the advancement in this area is really in details related to creating better explanatory variables.
For example, we provided you with some examples of detailed routing between micro-zones and
transit stops. How to better represent short trips and pedestrian and bicycle paths continues to be
an area of ongoing research, and a couple of agencies in San Francisco and Portland have
developed bicycle route choice models. Work on the Chicago ABM is focusing on incorporating
transit stop attributes in mode choice, building upon some previous work in Portland and for the
TCRP. These include attributes such as traveler information, presence of covered stops, fare
machines, safety\lighting of stops, and other nearby amenities).
The mobility attributes that we discussed in Webinar 7 are also very important, because they
have a strong conditioning effect on mode choices. This includes models that predict transit pass
and transponder holdings; bicycle ownership/usage; employer vehicle requirements and parking
subsidies; and persons with disabilities. In addition, there is some ongoing research on
representing the multi-dimensional choices of destination, mode and time of day in a unified
decision structure; however, thus far, these models have not been easy to implement in practical
898
models. There is also a lot of recent research on using space-time window concepts to constrain
choice sets for both destinations and modes and how to best implement that in a tour context.
Page 57
Trip mode
Actual mode used for each trip on tour
Constrained by tour mode
Rubber-banding
Out-of-direction travel cost for intermediate stops
Activity-Based Modeling: Tour Mode, Primary Destination,
Intermediate Stop Location and Trip Mode
57
In todays session, we covered tour mode choice models, trip mode choice models and
intermediate stop location choice models. A tour mode is the primary or preferred mode for the
tour. The tour mode ensures consistency between stop locations on the tour and between modes
used for trips on the tour. The tour mode choice model takes into account the round-trip level of
service between the origin and primary destination of the tour.
Trip mode is the actual mode used for each trip on the tour. Trip mode choice typically takes into
account one-way level of service between the trip origin and destination, but is constrained by
the tour mode.
Consistency is very important between choice dimensions, include tour mode and trip mode, the
location of intermediate stops on the tour with respect to tour origin and primary destination, and
899
the consistency between modes used and locations of stops. These consistencies are ensured
through constraints and situational variables.
The process of rubber banding refers to the measurement of the out-of-direction travel cost
incurred by stop locations on a tour.
900
Page 58
901
Page 59
February 2
February 23
March 15
April 5
April 26
May 16
June 7
June 28
July 19
August 9
August 30
September 20
59
902
Page 60
60
903
all aspects of accessibility. However, mode and destination decision are made on different time
scales, so a fully consistent model is not necessarily appropriate. Distance is often needed in
order to correctly measure how travelers perceive space when choosing a location, as opposed to
choosing a mode. Co-linearity between a mode choice log-sum and distance variables can be a
problem in model estimation. Often transformations of distance are used in estimation in order to
minimize the problem (log of distance, distance-squared, distance-cubed, etc.). Alternatively, one
might estimate a model with only a mode choice log-sum, and then constrain that parameter to
its estimated value and adjust distance terms in calibration in order to match the observed trip
length frequency distribution.
Are the outcomes from an activity-based model better for modeling managed lanes than in tripbased models?
Joel: We will cover this in more detail later, but the short answer is that yes, they are better. The
models provide more information for modeling managed lanes path choice, time-of-day choice,
mode choice, and even tertiary effects.
In the absence of a longitudinal survey, isnt it a bit risky to model tertiary effects? Is it reaching
too far?
John: Is it riskier to ignore them? We are relying upon cross-sectional data to inform choice
elasticities. However, it would be riskier to ignore the effects of variables on those choices.
While it would be ideal to have longitudinal data upon which to estimate the effects of changes
in inputs, this data is very difficult and expensive to collect. Therefore cross-sectional data is the
only data available, and we need to use it in order to capture important relationships. For
example, it is clear that sensitivities to level-of-service, land-use effects, etc. are important
determinants of auto ownership. Cross-sectional survey may not be a perfect data set but it is
appropriate and very necessary to measure the effect of such variables.
Has anyone modeled policies such as electric bicycles?
Joel: There is research exploring potential market penetration of electric bicycles and so forth,
but not any activity-based models that we are aware of that model such technologies explicitly.
However, one advantage of activity-based models is the use of models to do scenario-based
planning. In such exercises, one asserts the share or market penetration of a new technology,
such as electric vehicle penetration, and adjusts alternative-specific constants to reflect that
assertion. Then the model is run and outputs are summarized to determine the effects of that
policy on travel demand. The Metropolitan Transportation Commission has performed similar
analyses using their activity-based model for their Regional Transportation Plan and presented
their work at the 2012 Transportation Research Board Conference.
What are the key tour purposes used in activity-based models?
905
Joel: Mandatory work, school. Maintenance escort, shop, other maintenance. Discretionary
visiting, eating out, other discretionary. Additionally, tours are classified as to whether they are
home-based or work-based (according to their origin or anchor location). Finally, some models
explicitly model joint travel; therefore, another classification is given to fully-joint or partiallyjoint tours. See webinars 3, 4, and 8 for more details.
906
907
Page 1
Activity-Based Modeling
Session 11: Network Integration
908
Page 2
Acknowledgments
This presentation was prepared through the collaborative efforts
of Resource Systems Group, Inc. and Parsons Brinckerhoff.
Presenters
Joe Castiglione and Peter Vovsha
Moderator
John Gliebe
Media Production
Bhargava Sana
Activity-Based Modeling: Network Integration
Resource Systems Group and Parsons Brinckerhoff have developed these webinars
collaboratively, and we will be presenting each webinar together. Here is a list of the persons
involved in producing todays session.
Joe Castiglione and Peter Vovsha are co-presenters. They were also primarily responsible
for preparing the material presented in this session.
John Gliebe is the session moderator.
Content development was also provided by John Gliebe, Jason Chen, Joel Freedman, and
Rosella Picado. John Bowman and Mark Bradley provided review.
Bhargava Sana was responsible for media production, including setting up and managing
the webinar presentation
909
Page 3
February 2
February 23
March 15
April 5
April 26
May 17
June 7
June 28
July 19
August 9
August 30
September 20
3
For your reference, here is a list of all of the webinars topics and dates that have been planned.
As you can see, we are presenting a different webinar every three weeks. Three weeks ago, we
covered the tenth topic in the seriesTour and Trip Mode, Intermediate Stop Location. This
session covered some of the key mode and destination choice components of the activity-based
model system.
Todays session is the eighth of nine technical webinars, where we will cover the details of
activity-based model design and implementation. In todays session, we will describe how
different activity-based model systems are integrated with network or supply models, and key
considerations of this linkage.
910
Page 4
Learning Outcomes
911
Page 5
There are two senses in which we can use the term network integration First, the term can
refer to a condition or an outcome that is achieved when the network-derived level of service
variables used to predict activity generation, destination, mode, and trip timing decisions are
consistent with the level of service that results when these trips are loaded onto networks during
the network assignment steps. Second, the term can refer to the data structures and procedures
that are used to achieve this outcome.
912
Page 6
Outline
Basic terminology
Why network integration is important
Where network integration fits into travel model
systems
Theory and model formulation
Data sources
Benefits and costs of network integration
Ongoing research
Questions and Answers
Activity-Based Modeling: Network Integration
In this webinar, we will consider why network integration is important, and how it functions
within the overall travel model system. Some basic theories and formulations and associated data
requirements for implementing different types of network integration will be discussed, and
consider the tradeoffs associated with different approaches. Finally, we will cover ongoing
research into network integration and leave time for questions and answers.
913
Page 7
Terminology
Demand Models
Supply Models
Feedback
Convergence
Equilibrium
This slide shows basic terminology for that will be used in this session and other sessions.
Demand Models: Tools used to generate estimates of the type, amount, locations, mode
and timing of the demand for travel. Typically, this refers to the dimensions of travel that
are predicted by the first three steps of a traditional 4-step model (as distinct from the
final assignment step) or that predicted by an activity-based demand model. Demand
models can be basic or extremely complex.
Supply Models: Tools used to generate estimates network performance measures (such
as link flows and congested travel times) which are used as key inputs to demand models.
Like demand models, supply models can be quite basic in their formulation or
significantly more complex.
Feedback: Refers to the process through which information generated lower in the
model system (such as congested travel times from network assignment) is used as direct
or indirect input the models higher in the model system (such as activity generation).
914
The purpose of feedback is to ensure that the final model outputs are consistent with the
model system inputs and assumptions.
Convergence: The condition when the impedances or level-of-service measurements
used as the basis for accessibility measures and as key inputs to the destination and mode
choice models are approximately equal to the travel times and costs produced by the final
network assignment process. Convergence is necessary in order to ensure the behavioral
integrity of the model system, as is considered both with the context of the network
assignment process, as well the overall model system.
Equilibrium: Equilibrium typically refers to the condition where during the network
assignment no traveler can decrease travel effort by shifting to a new path. This is known
as user equilibrium, although other conditions such as system optimum can also be
pursued.
915
Page 8
Demand Models
Predict dimensions of travel demand (activity generation,
destination, mode)
Comprised of linked demand model components
May be applied at aggregate (zones) or disaggregate levels
(persons, HHs)
Transportation supply availability and network performance
variables derived from supply model may appear in the utility
expressions of any of these components
Accessibility variables (simplified log-sums) typically used to
represent complex hierarchical travel choices
Model components interact, causing second-order effects on
model components that do not use network variables directly
Activity-Based Modeling: Network Integration
Demand models are tools used to generate estimates of the type, amount, locations, mode and
timing of the demand for travel. Typically, this refers to the dimensions of travel that are
predicted by the first three steps of a traditional 4-step model (as distinct from the final
assignment step) or that predicted by an activity-based demand model. Demand models can be
extremely basic (such a simple cross-classification trip generation model) or extremely complex
(such as an intra-household activity generation/coordination model). There are usually a series of
individual model sub-components that we refer to collectively as a single demand model for
example, the trip or activity generation model component is distinct from the destination choice
or distribution model component, which is distinct from the mode choice model component.
These and other components are executed in sequence.
Demand models can be applied at an aggregate level, such as zones, as is in most traditional tripbased models, or may be applied at the level of individual persons or households, as is the case in
activity-based models. Critical inputs to demand models are measures of transportation supply
availability and transportation system performance that are derived from the supply model. In
916
some cases, these measures are used directly in the demand model components, while in other
cases the measures may be incorporated into accessibility measures and used indirectly in
demand model components.
917
Page 9
Tools used to generate estimates network performance (such as link flows and congested travel
times), which are used as key inputs to demand models. Supply models can be basic (such a
simple aggregate static equilibrium model) or more complex (such as a DTA or traffic microsimulation model that incorporates detailed operation attributes such as signal timing). Supply
models require as input information about travel demand, such as the locations, timing, and
modes used for travel. These estimates of demand are applied to representations of network
structures comprised at minimum of links and nodes but sometimes incorporating additional
network attributes which are used to predict the paths through the network that will be used to
satisfy this demand. In traditional static assignment supply models, mathematical functions are
used to estimate congested travel times given supply and demand inputs, although more recent
supply modeling techniques rely less on these volume delay functions. Because travelers
choices are influenced not only by travel times but also by monetary costs, supply models should
be configured to convert these costs into travel times so that they can be incorporated into the
path-building procedure.
918
Page 10
Feedback
Use of outputs from a later (lower) model
component as input into an earlier (higher) model
component
Intended to ensure that the final model outputs are
consistent with the model system inputs and
assumptions
Extent of feedback and equilibration rules relate to
structure of model
Necessary in both traditional trip-based models as well
as activity-based models
Activity-Based Modeling: Network Integration
10
Feedback refers to the process through which information generated lower in the model system
(such as congested travel times from network assignment) is used as direct or indirect input the
models higher in the model system (such as activity generation). Feedback is important in both
traditional trip-based models as well as activity-based models in order to ensure that the final
model outputs are consistent with the model system inputs and assumptions. The exact nature of
this feedback is related to the structure of the model. For example, if in an activity-based model
system the activity generation component uses accessibility measures that reflect network
performance then the supply model outputs should be fed back to update these accessibility
measures before running any subsequent model components. However, if a traditional trip-based
model system incorporates no network performance measures or network-based accessibility
measures in trip generation or trip distribution, then it may only be necessary to feed-back
network performance information through the mode choice step.
919
Page 11
11
Model convergence is necessary to ensure the behavioral integrity of the model system, and to
ensure that the results will be useful in a policy context. The network performance or level-ofservice measurements used as the basis for accessibility measures and as key inputs to demand
model components must be approximately equal to the travel times and costs produced by the
final network assignment process. In a travel model system, there are at least two types of
convergence that we need to consider: network convergence and system convergence. When we
talk about convergence, we are implicitly talking about convergence to something. Typically
this means for networks that we are converging to an equilibrium condition (usually a
deterministic user equilibrium where, for each time period-origin-destination combination all
used routes have equal travel times, and no unused route has a lower travel time). For the overall
model system, this usually means that we are converging to a stable solution (rather than an
optimal solution as in the network context). It should be noted that in the context of an integrated
demand and network simulation model system, an essential precondition for pursuing overall
model system convergence is establishing network assignment convergence.
920
Page 12
Behavioral
Demand patterns produce supply cost
Supply costs influence demand patterns
Structural
Demand models results are input to supply model
Supply model results are input to demand model
Practical
Policy/investment choices must be informed by stable, repeatable results
Exchange of consistent information required to produce stable,
repeatable results
12
Page 13
Influenced by
Demand model methods
Supply model methods
Model integration methods
Activity-Based Modeling: Network Integration
13
Travel demand forecasting model systems are tools with which we try to measure the impacts
(on travel choices, on network performance) of different policy and investments. It is essential
that they generate the same outputs when fed the same inputs, within some acceptable range. The
tool would be useless if it could not generate reproducible results. Activity-based model systems,
which are typically implemented using Monte Carlo simulation techniques may produce slight
variations in outcomes because their probabilistic nature, although as prior webinars have
described most activity-based models incorporate features that significantly attenuate these
stochastic effects. Ultimately, we want to avoid using a model system that generates multiple
outputs that are sufficiently different that they may lead to different decisions.
Not only is it important that models produce repeatable, stable and dependable results, but also
that when models are used to compare alternative scenarios the differences in the model outputs
reflect difference in the input assumptions or parameters, and are not attributable to model error.
If model output differences reflect issues with the model implementation rather than difference
in the scenarios, the tool will not be useful. In addition, we want to avoid using a model system
922
that is dependent on the specifics of the network starting conditions, as this may thwart efforts to
produce repeatable results. For example, we want to our model system to produce similar final
results regardless of the seed impedances that we may use in the model systems initial
iteration. We also want to avoid using a model system where the results oscillate between
multiple outcomes in ways that may be consequential to decision making.
Our ability to produce repeatable and stable results is influenced by the resolution of the methods
used within both the demand and supply models, but is perhaps most crucially affected by the
methods we use to integrate our demand and supply models.
923
Page 14
14
When travel demand forecasting first emerged, the questions that practitioners asked of travel
models were simpler, such as how to size a given facility given the expected locations of future
populations and jobs. These days, decision-makers rely on travel demand models for answers to
more complex policy questions, and expect that these models are appropriately sensitive to the
complex behavioral responses.
For example, adding capacity to a roadway segment may not only result in diversion of traffic as
people seek reduced travel times, but might release latent demand that was suppressed due to
exiting congestion. Similarly, a pricing scenario might influence not only the use of specific
routes, but also timing and mode choices, destination choices, and even the generation of
activities.
And the complex policy questions are not strictly limited to transportation investments.
Decision-makers want travel models that are appropriately sensitive to the effects of compact,
mixed use, and transit-oriented land use. They want models that can help inform travel demand
924
management strategies such as flexible work schedules. And (ideally) they want models that can
tell them who is impacted by these transportation and land use policy and investment decisions.
The complex questions necessitate models that appropriately and consistently capture the
relationship between travel demand and supply. If network pricing is to change by time-of-day,
then ideally network information (times, costs) at a time resolution consistent with this pricing
scenario can be fed back from supply model to demand model. Similarly, if value-of-time
distribution information is used when predicting demand, then ideally this segmentation would
be reflected in the configuration of the supply model.
925
Page 15
Examples to avoid
many people coming home really late or working very
short days
transferring 3 times or walking long distances after
disembarking from transit
leaving young children stranded at school or home alone
15
Related to the notion maintaining a consistent representation of choices between the demand and
network supply components of the model system is the issue of respecting temporal and spatial
constraints. Activity-based models contain intrinsic logic that already constrains individual travel
choices. For example, members of household that dont own cars typically arent allowed to use
drive alone modes. People dont drive home from work alone if they havent driven to their
work location. Travelers arent allowed to depart from a location that wasnt the destination of
the prior trip.
Ideally, the supply component of the model system can reflect these constraints. For example,
transit skims should have sufficient temporal detail that they can reflect the fact that after a
certain time transit service is significantly less frequent, and thus shouldnt be considered as an
alternative. The transit network processing should discourage extreme behavior such as making
three or more transfers, or assuming long-walk egress.
926
Maintaining budgets and constraints is not usually considered one of the primary challenges in
integrating activity-based demand models with current network supply models. However, as
network supply models incorporate ever increasing levels of temporal, spatial and behavioral
detail (such as the linked nature of trips on a tour), these issues may become more pronounced.
927
Page 16
Build Alternative(s)
16
To understand the impacts of network integration, lets revisit the example bridge expansion
transportation planning and policy project.
For this scenario analysis, we will be considering a number of alternatives: a no-build alternative
and a various configurations of the build alternative. In the no-build alternative the bridge has 4
lanes (2 in each direction), there are no tolls, and the transit fare stays the same all day. In the
various build alternatives, there are 6 lanes on the bridge. In some alternatives the two additional
lanes will be HOV lanes all day, while in other alternatives the two additional lanes will be HOV
lanes only during peak periods. In addition, in some build alternatives there will be a new toll
that is the same across the entire day, while in other build alternatives there will be a toll that will
be only applied during peak periods, or when certain levels of congestion occur. Finally, in the
build alternatives regional transit fares will be higher during peak periods.
How would the analysis of these alternatives be impacted by different network integration
schemes?
928
Page 17
Previous presentations have described the potential long-term and medium-term effects of the
bridge expansion, such as changes in the usual work and school locations, changes in levels of
vehicle ownership, and the types of vehicles owned, and changes in the transit pass or toll
transponder adoption.
929
Page 18
18
In addition to changes in these medium and long term mobility choices, travelers may also
have short term responses to the bridge alternatives. For example, in response to the different
levels of congestion and costs associated with the different alternatives, travelers may choose
new locations for discretionary purposes such as shopping or personal business potentially
avoiding bridge crossings if subject to an additional toll or conversely choosing to make a bridge
crossing to take advantage of congestion relief associated with using HOV lanes or provided by
congestion-based tolling. If tolls, transit fares, or the availability of HOV lanes varies by time-ofday, then travelers may choose to change the timing of their travel to either take advantage of or
to avoid differences in travel time and costs by time-of-day. Similarly, these differences in travel
times and costs may induce some travelers to choose new routes in some cases this may mean
new travelers using the bridge, while in other cases it may mean existing bridge travelers
selecting alternative routes. The ability of the model system to be sensitive to these potential
traveler responses depends upon the how the network supply components of the model system
are integrated with the demand components.
930
Page 19
Supply model
Roadway and transit network coding by time-of-day to reflect changes in modal availability
and costs, and addressing key issues such as directionality by TOD
Assignment and skim processes that reflects variations in times and costs by time-of-day,
segmentation by mode (SOV/HOV, toll/free), and market (VOT class)
Integration
Data exchange
Feedback / iteration
Activity-Based Modeling: Network Integration
19
In order to ensure that the model is appropriately sensitive to the various alternative
configurations, careful consideration must be given to the design of the demand model
components, the supply/network model components, and the integration between the two.
On the demand side, ideally the model incorporates accessibility measures which influence
longand medium-term choices as well as activity generation and these measures are sensitive to
changes in accessibility by time-of-day and reflective of all modes. Given that a fundamental
aspect of the project involves variations in mode availability, tolls, and fares by time-of-day, the
demand model should incorporate a temporal resolution that is fine-grained enough to represent
the policies as well as the changes in network performance by time-of-day provided by the
network supply model. The demand model must also be able to distinguish between SOV and
HOV alternatives given the different networks (and by extension, different network performance
measures) associated with these alternatives, and might optionally include toll and notoll
alternatives. It is also critical that the demand model incorporate a behavioral resolution that can
931
reflect different sensitivities to congestion, tolls and fares depending on the traveler, travel
purpose and other travel attributes.
On the supply side, the roadway and transit networks should incorporate information about
modal availability (such as HOV lanes) as well as about how the network configurations and
costs change by time of day, consistent with the policies to be evaluated will tolls or fares vary
only by broad multi-hour time period, or will they vary by finer time periods such as individual
hours? A key aspect of this is directionality, especially for transit transit networks need to be
coded to reflect the true level of service provided by direction. In order to exploit the modal and
time period information coded in the networks and provide relevant information to the demand
model, it is also critical to ensure that the network assignment and network skim methods reflect
the variations in times and costs by time-of-day, mode and potentially other dimensions such as
value of time.
Finally, the core issues of integration also need to be considered what information is being
exchanged between the demand and supply components, and how are the demand and supply
components interacting in an iterative feedback framework in order to ensure consistent and
reasonable results? The demand model must provide information about travel demand that is
sufficiently segmented by mode, time-of-day, traveler class and other attributes, while the
supply model must provide information about network performance that is similarly segmented.
932
Page 20
Supply Models
Highway and transit assignment, traffic simulation by time
periods
Integration / Connectors
Feedback loops, convergence monitoring
20
As the bridge scenario discussion illustrated, when considering how to achieve network
integration, we need to consider three elements of the model system:
The demand model which generates activities and predicts the location and timing of
these activities as well as the mode of transport and which provides the required
information to the supply model
The supply model which assigns the demand generated by the activity-based demand
model to roadway and transit networks using either static of dynamic assignment
methods and which generates measures of transportation system performance or
impedance for input to the activity-based demand model system; and
The connectors which enable the feedback loops and handshake between the demand
and supply components, and which may also assess convergence.
933
Page 21
Demand Model
Supply Model
Feedback Process
NO
Estimates of
demand by
location, time
period, mode,
etc.
Estimates of
network
performance by
location, time
period, mode, etc.
YES
END
Activity-Based Modeling: Network Integration
21
This figure illustrates the relationships amongst the three types of components in the integrated
model system. Note that this figure is relevant regardless of whether the demand model is a more
typical aggregate, trip-based model or a disaggregate activity based model. It illustrates that the
demand model uses travel time and cost information in the form of skims, in addition to other
attributes associated with travelers (such as income) and locations (such as the amount of
employment) to predict activity-travel events.
The network supply models then apply this demand to roadway and transit networks to estimate
volumes and associated times and costs. This new time and cost information is then used to
develop updated skims. These new skims are then used to rerun the demand model to predict a
revised set of activity-travel events and again this demand is assigned to roadway and transit
networks to develop revised estimates of volumes, times and costs. The convergence of the
model system to a stable solution is assessed, typically using measures that consider changes in
the demand flows by geography or by changes in the skims, and if a pre-specified threshold is
met, the process terminates.
934
If the threshold is not met, then the demand and network supply models are run again,
convergence is checked, and the process repeats until either the threshold is met or the system
reaches a maximum number of iterations.
935
Page 22
Activity generation
Tour and stop location
Tour and trip mode
Time-of-day
22
Activity-based model systems incorporate significantly more choice dimensions than traditional
4-step models. These additional dimensions are reflected both in the number of components that
comprise the model system, the complexity of these components, and the type and amount of
detailed information that is exchanged via the model linkages between the demand and supply
components
While there are a number of similarities between trip-based and activity-based models systems
(for example, both types of models include mode choice components), activity-based model
systems typically incorporate more model system components. As prior webinars have
illustrated, some of these additional components provide more detailed sensitivity to choice
dimensions, such incorporating distinct models for predicting the primary destinations for tours
as well as for predicting the likely locations for intermediate stops on these tours. Other
components provide sensitivities to choice dimensions that are often not embedded within a
traditional trip-based model system, such as explicit models of time-of-day.
936
Activity-based model components are also typically more complex than traditional trip-based
model systems. For example, AB model components may explicitly incorporate information
about constraints, such as the time windows available to individuals to participate in activities, or
the maximum distance than can be travelled within an available time window. And of course,
activity-based models systems employ higher levels of behavioral, temporal and even spatial
detail, using individual persons and householders as decision-makers, representing time in small
time slices such as one-hour, half-hour, or even continuously.
The fine-grained behavioral, temporal, and spatial resolution of activity-based demand models
and the complexity of the models that exploit this detail require that significant consideration be
given to the coding the linkages between demand model system components. There are two
primary linkages in an activity-based model system: from the demand model to the network
supply model, and from the network supply model to the demand model.
937
Page 23
Trip-Based Model
Long-term choices
Trip generation
Trip distribution
Mobility choices
(transit pass ownership)
Time-of-day factoring
skims
Mode choice
Activity generation
Destination choices
Mode choice
Time-of-day choices
Skims/logsums
Network assignment
Network assignment
Activity-Based Modeling: Network Integration
23
This slide shows parallel structures of trip-based and activity-based demand-supply linkages. The
primary components of a traditional trip-based model system generation, distribution, mode
choice have analogs in the activity-based model system. The first of the two primary linkages
in the model system that we will consider is the linkage from the demand model to the network
supply model. As in a trip-based model system, the primary information that is being conveyed
are estimates of travel demand. Careful consideration needs to be given to the type and format of
how this travel demand information is conveyed.
938
Page 24
Market Segmentation
Overall travel market is comprised of submarkets
Activity-based models provide more flexibility and
efficiency in handling market segmentation
Submarkets are differentiated by key attributes, such as:
Mode
Time-period
Value-of-time
Trip Rate
Medium
Income
Household
Low Income
Household
0 Vehicle
Drive alone
Vehicle Less
than Workers
Shared Ride
Vehicle more
than Workers
Walk to Bus
0 Vehicle
High Income
Household
24
Let us first consider the issue of market segmentation. Market segmentation refers to the
treatment of the overall travel market as comprised of a series of smaller markets that are
differentiated by some key attributes. For example, referring back to our bridge expansion
example, we can consider the transit segment (those travelers who chose to use transit) as distinct
from the auto segment (those travelers who chose to drive alone or share rides). In our network
supply model we want to treat these two different market segments separately we allow transit
users to find the best transit path available to them and also prevent them from driving.
Conversely, we want to make sure that travelers who chose to drive alone or share rides dont
end up using transit instead. Note that we have assumed in this example that our demand model
determines the mode (or sub-mode) that a traveler chooses.
Market segmentation applies not only to modal choices, however, but may also extend to other
choice dimensions as well, such as time-of-day, value-of-time, or purpose. From a network
integration perspective, one of the key advantages of an activity-based model framework is that
it allows tremendous flexibility in how we define market segments both within the demand
939
model, within the network supply model, and in the linkages between these two components. In a
trip-based model system we have less flexibility to define market segments, primarily due to the
fact that as we increase the number of market segments there is a combinatorial effect that may
lead to the proliferation of a huge number of segments, each of which may require the
maintenance of multiple matrices regardless of how big the market segment truly is.
940
Page 25
Trip tables not produced until assignment step, more flexibly specified
Trip lists may be converted to trip tables by aggregating over multiple dimensions
Time of day
Activity types
Person types
occupancy class
Transit submode /access mode
Other EZ pass, transit and parking pass holders, etc.
25
In contrast to the proliferation of matrices and all the associated computation and storage
challenges that results from increased market segmentation in a traditional matrix-based and tripbased model system, activity-based model systems support much more flexible market
segmentation due to the list-based nature of the activity-based demand model simulation. One
of the outputs from the activity-based demand model is a list of trips that contains all the detailed
spatial, temporal, behavioral, and socio-demographic information associated with the trip and
traveler.
When assigning travel, most network supply models require as input a set of origin-destination
matrices segmented by mode and time-of-day. The detailed information contained in the listbased activity-based demand model can be aggregated to virtually any market segmentation. For
example, if a user wanted to transition from using a three hour peak period assignment to 3
separate 1-hour peak hour assignments or to assign by value of time class, they only need to
revise the aggregation process and make associated changes to the network supply model
assignment and skim scripts.
941
Page 26
Auto
Drive
alone
Shared
ride 2
Shared
ride 3+
Walk(9)
Transit
Walk
access
PNR
access
School
Bus(26)
KNR
access
Local
bus(11)
Local
bus(16)
Local
bus(21)
HOV(7)
Express
bus(12)
Express
bus(17)
Express
bus(22)
Pay(8)
BRT(13)
BRT(18)
BRT(23)
LRT(14)
LRT(19)
LRT(24)
Commuter
rail(15)
Commuter
rail(20)
Commuter
rail(25)
GP(1)
GP(3)
GP(6)
Pay(2)
HOV(4)
Pay(5)
Bike(10)
26
This figure provides an example of the segmentation of the auto market. In this scheme, the auto
travel market is subdivided into a total of 8 market segments that correspond to distinct
alternatives:
We treat these as separate choices in the demand model because they are associated with
different time and cost measures. The drive alone pay alternative may have a higher monetary
costs associated with it, but may have lower travel times. Depending on the individual traveler
and decision making context, a travel may choose to select of these modes.
942
In establishing the integration with the network supply model, we want to consider this as a
separate market segment because each choice may be subject to different opportunities or
constraints. For example, if the mode chose for a given trip is drive alone-pay then when we
assign this using the network supply model we can allow the trip to use pay/toll facilities as well
as general purpose lanes. However, we also want to restrict this trip from using HOV lanes.
943
Page 27
Nonmotorized
Auto
Drive alone
Shared ride
2
Shared ride
3+
Walk(9)
GP(1)
GP(3)
GP(6)
Bike(10)
Pay(2)
HOV(4)
Pay(5)
Transit
Walk
access
PNR
access
School
Bus(26)
KNR
access
Local
bus(11)
Local
bus(16)
Local
bus(21)
HOV(7)
Express
bus(12)
Express
bus(17)
Express
bus(22)
Pay(8)
BRT(13)
BRT(18)
BRT(23)
LRT(14)
LRT(19)
LRT(24)
Commuter
rail(15)
Commuter
rail(20)
Commuter
rail(25)
27
This figure provides an example of the segmentation of the transit market. In this scheme, the
transit travel market is subdivided into a total of 15 market segments or alternatives. These 15
market segments represent the combination of 5 primary transit modes (local bus, express bus,
BRT, LRT, and commuter rail) and 3 transit access modes (walk, park-and-ride, and kiss-andride). Again we treat these as separate choices in the demand model because they are associated
with different time and cost measures, and potentially availability.
In establishing the integration with the network supply model, we want to consider each of these
as a separate market segment when assigning demand and generating skims because each choice
may be subject to different availability or constraints. For example, if the mode chosen for a
given trip is walk access-LRT then when we assign this using the network supply model we
can allow the trip to use LRT routes, and potentially other routes that are used to access LRT,
such as local bus; however, we probably also want to restrict this trip from using commuter rail
to reflect the hierarchy of transit sub-modes typically employed in representing transit services.
944
Page 28
945
Page 29
Mobility choices
(transit pass ownership)
Trip distribution
Destination choices
Mode choice
Mode choice
Time-of-day factoring
Time-of-day choices
Network assignment
Network assignment
Skims/logsums
Activity generation
skims
Trip generation
29
This slide shows parallel structures of trip-based and activity-based supply-demand linkages. The
second key linkage in the activity-based model system that we will consider is the linkage from
the network supply model back to the demand model.
As in a trip-based model system, the primary information that is being conveyed are estimates of
network performance such as travel times and costs, often referred to as network skims. This
feedback of skims is critical to achieving the converged or stable model results that are necessary
for the model to be useful as an analytic tool. As with the demand model to network supply
model linkage, careful consideration needs to be given to how this network performance
information is fed back and incorporated into the activity-based demand model component.
946
Page 30
30
Network performance measures (or level of service variables) as well as accessibility variables
that are calculated using these measures in combination with land use attributes, appear in
virtually all subcomponents of activity-based travel demand models. For example:
948
Page 31
No non-motorized assignment
Activity-Based Modeling: Network Integration
31
Trip-based models often employ a much coarser representation of network performance by time
of day. For example, in smaller urban areas it is not uncommon to perform a single peak period
and a single off-peak period assignment for highway, to perform only limited transit assignments
(such as the AM peak only), and to provide no non-motorized assignment. In larger urban areas,
trip-based models may incorporate more a few, such as both and AM and PM peak period
assignment, distinct midday and evening assignments, and to perform transit assignments using
these same resolutions.
949
Page 32
32
Market segmentation does not simply apply to the grouping of demand by modes of travel. One
of the primary other types of market segmentation used in both traditional trip-based models as
well as within activity-based models is segmentation by time-of-day. And this market
segmentation affects the design not only of how demand information is transmitted to the
network assignment and skimming models, but also how network skim information is
transmitted back to the activity-based demand model.
Note that this discussion considers segmentation by time-of-day in the context of the demand
model-network supply model integration. Specifically, what time periods are used when demand
is assigned using the network supply model, and what time periods are used when skimming
network performance measures for input to the demand model? This segmentation is distinct
from, but related to, the temporal resolution of the demand models themselves that is, most
activity-based demand models function using a temporal resolution that is much finer grained
(hours, half hours, or even finer) than the temporal resolution of the network performance
950
indicators input to the models. Achieving consistency in temporal resolution across demand and
supply models is a key current research topic.
Ideally, in order to have the greatest sensitivity in the integrated demand-network supply model
system, we would input to the demand model information about network performance generated
by the supply model that is consistent with the temporal resolution of the demand model, and we
would assign this demand to model networks using this same temporal resolution. However,
theoretical and practical concerns necessitate simplification the runtimes and hardware
requirements associated with generating, storing, and accessing network skims for detailed time
periods quickly become onerous, and the using static assignment methods for short time periods
may be problematic. Some of these issues may be addressed by using DTA, which we will
discuss later in this presentation.
Practically, we need to consider a few issues:
Our assignment time periods and assignment skim periods should be consistent;
More detailed assignment and skim periods provide better model system sensitivity,
particularly to changes by time of day and mode;
Time period definitions should reflect potential policy applications it wont be possible
to test the impacts of hourly changes in tolls, fares, reversible lanes unless information at
this level of temporal detail can be generated by the network supply model for input to
the demand model; and
Time period definitions should reflect the regional context larger regions with more
complex transportation systems may be more subject to phenomena such as peak
spreading, or may have more diverse modal alternatives with service differential by
detailed time of day.
951
Page 33
Trip distribution
skims
Mode choice
Time-of-day factoring
Network assignment
33
Most typical 4-step travel model systems do not incorporate explicit time-of-day models. Rather,
a set of fixed factors (potentially segmented by purpose and mode) are applied post-mode choice
that transform daily production-attraction format trip matrices into origin-destination trip
matrices by time period.
Applying time-of-day factors to the trip tables at this point in the model system is a
simplification that makes it possible to generate assignment results by time-of-day, but without
requiring the proliferation of matrices that would result from incorporating time of day choice
models, or even from applying fixed time of day factors earlier in the model stream.
However, the use of time-of-day factors represents a significant compromise in the sensitivity of
the model system. For example, the model would be unable to respond to the time-of-day
changes that might result from the expansion or reduction of capacity in a congested corridor.
The model would also not be sensitive to the fact that as congesting increases, travelers start to
use the shoulders of the peak known as peak spreading. And perhaps most significantly,
952
there is no information on how these factors should change in the future given more demand. In
this last example, the use of fixed base year factors in the future year might significantly overestimate the amount of congestion encountered by travelers, and thus lead to unrealistic
responses by other model system components.
Some trip-based models do incorporate peak spreading models that begin to capture some of the
sensitivity of travelers to changing the timing of their trips. However, such features are not
typical components of 4-step trip-based model.
953
Page 34
Mobility choices
(transit pass ownership)
Activity generation
Destination choices
Time-of-day choices
Skims/logsums
Mode choice
Network assignment
Activity-Based Modeling: Network Integration
34
In contrast to trip-based models, most activity-based models employ more detailed time period
information when integrating the demand and network supply components of the model system.
Some of the earliest activity-based models systems implemented in the United States used
relatively aggregate time periods for skimming and assignment, even when the demand models
operating using quite detailed time-of-day models.
954
Page 35
CMAP
8 time periods used in assignment and skimming (temporal resolution of
demand component is half-hour)
Detailed auto and transit sub-modes
SACOG
12 time periods in skimming and assignment (temporal resolution of
model is half-hour)
Detailed auto and transit submodes
Network skimming and assignment at zone level, enhanced with parcellevel geographic information
Activity-Based Modeling: Network Integration
35
In contrast to trip-based models, most activity-based models employ more detailed time period
information when integrating the demand and network supply components of the model system.
Some of the earliest activity-based models systems implemented in the United States used
relatively aggregate time periods for skimming and assignment, even when the demand models
operating using quite detailed time-of-day models.
For example, the MORPC demand model uses 1-hour time periods when generating and
scheduling activities, but only employs two time periods when feeding back skim information to
the demand model. In contrast, the SFCTA integrated model system uses a consistent set of time
periods for generating and scheduling activities, and for network assignment and skimming, but
these time periods are much broader than one hour. The recent updates to the SACOG model
have resulted in an implementation in which the demand model uses half-hour time periods for
generating and scheduling activities, and uses 12 time periods for network assignment and
skimming (hourly during the peaks, and broader time periods during the midday and
evening/night).
955
All of the models described above incorporated fairly detailed auto and transit sub-model
alternatives in the demand model, and an analogous market segmentation in the demand and
network supply model linkages. In a sense, we can also consider the representation of space in
the model system as a type of geographic segmentation.
956
Page 36
36
In general, there is a pronounced move towards incorporating greater levels of temporal detail in
both roadway and transit assignment and skimming, as evidenced by recent work in Chicago,
New York, Sacramento, and other regions.
In addition to the detailed segmentation by time of day, some advanced activity-based demand
and network supply model integration efforts are incorporating more advanced assignment and
skimming approaches that included non-motorized modes such as bikes, or that include roadway
assignments and skimming that incorporate value-of-time segmentation to capture different time
and cost tradeoffs.
957
Page 37
Nontoll
SOV
Non-toll
HOV2
Non-toll
HOV3+
Toll
SOV
Toll HOV2
Toll
HOV3+
Auto + external +
airport low VOT
Auto + external +
airport high VOT
11
10
12
Commercial
13
14
Light truck
15
16
Medium truck
17
18
Heavy truck
19
20
37
This is an example of a multi-class assignment setting use for the Chicago activity-based model.
After addressing vehicle types, occupancy categories, and value of time categories at least for the
most important classes it results in 20 classes. It is difficult and computationally onerous to add
more details in this aggregate setting. Individual micro-simulation DTA is a long-term solution.
958
Page 38
38
An interesting implementation issue relates to differences in the way that activity-based models
access and use skims relative to trip-based models. First, trip based models do not typically use
skims as comprehensively throughout the model system as activity-based models. Of course,
mode choice models use skims, and distribution models typically do as well. In the rare instances
where a trip-based model incorporates a time-of-day component then this model will also use
skims. But trip-based models are implemented within a matrix framework, looping first on
origins and then over destinations. This approach allows for efficient access to skim values for
large batches of trips in a single operation.
In contrast, activity-based models incorporate skim information throughout virtually all
components of the entire model system destination choice, mode choice, time of day choice,
and in the generation of accessibility measures used as input to activity generation. Due to the
agent-based micro-simulation framework in which the activity-based model is applied, random
access of skims during the simulation is required. This is computationally challenging, and
necessitates much greater memory requirements, and efficient means of retrieving the skim data.
959
Page 39
39
The input network supply data needed to integrate with an activity-based model system are not
significantly different than required to integrate with a traditional trip based model, but more
temporal and modal detail is typically required. A key advantage of activity-based models is that
they provide a framework for the consistent treatment of time and space and reflect realistic
availability constraints. To maintain consistency with and provide good information to the
activity-based demand model, the network supply model should ideally also incorporate detailed
information on network supply, particularly by time-of-day. This should include information on
changes in capacity by time-of-day, such as the presence of HOV and reversible lanes, and
transit service headways fares and coverage. Coding and maintaining information about roadway
supply by time of day is relatively straightforward as this information can usually be simply
coded as an attribute on a common base network. Coding and maintaining information on transit
supply by time-of-day can be considerably more onerous due to the significant numbers of
variations in transit route alignments and service levels. For example, a single transit route may
have a number of patterns with different termini and service frequencies. In some cases, agencies
960
have developed detailed transit network by time of day, while in other instances simplifying
assumptions are made, such as PM transit service as a transpose of AM service.
961
Page 40
40
This picture shows a small section of a highway network model in link-node format. This should
be familiar to most modelers out there. The gray box on the right lists important attributes of
links such as:
962
Page 41
zone/centroid
connectors
41
This slide shows a small section of hyper (highway/transit) network. Shown are links, nodes,
zone/centroid connectors, and transit access links. Transit routes are highlighted by color based
on different routes (turquoise for an express route following the freeway; red and green for local
routes on surface streets). This should be familiar to most travel modelers.
963
Page 42
Counts, screenlines
VMT / VHT
Speeds, travel times archived ITS data
Transit boardings / alightings
Transit ridership by line
42
Weve been discussing the network skims as a critical output of the network supply model
because it is the information contained within these skims that make the model system truly
integrated, where each component is informing the other. However, most practitioners are also
interested in other key outputs from the network supply model, link and segment-based volumes
and speeds. To assess how well both the network supply model as well as the overall demandsupply model system are performing, it is critical to have a robust model validation dataset. In
general, the types of measures used to validate an integrated activity-based travel demand model
systems are very similar to those used to validate a traditional trip-based models: a distribution
of counts and speeds across different locations, facility types and vehicle classes, as well as
transit boardings, alightings, and loads. The most significant difference is that these validation
data are ideally available at a level of temporal detail that are consistent with the time periods
used in the network assignment model. This may presents a minor additional burden as the time
period detail is greater than that used in most trip-based models. In addition, there are also some
validation data related to model components unique to the activity-based model system, such as
park-and-ride lot utilization.
964
Page 43
User Equilibrium
Static definition
Implies convergence criteria relative gap
Analogous to what we want to achieve at a system level
Activity-Based Modeling: Network Integration
43
Given the detailed information that is required to support the integration of the activity-based
demand model and the network supply model, there are a number of special issues that must be
considered. For example, in addition to the more detailed transit sub-modal detail that is usually
embedded in activity-based model systems, the tour-based nature of activity-based models and
their more comprehensive treatment of the temporal dimension necessitates the explicit treatment
of the directionality of transit network coding. For example, transit drive access connectors must
be bidirectional in order to represent the fact that people from park-and-ride lots back to their
homes after work most network software builds one-directional drive access links. Also related
to park-and-ride, the explicit representation of parking lot capacities supports a more realistic
treatment of drive access transit.
In addition, careful consideration should be given to the coding of demand access points in travel
model system networks. While some activity-based model systems have been implemented using
traditional travel analysis zones (or TAZs), many recent AB model development efforts have
used finer grained representations of space, such as subzones that are similar to census blocks, or
965
actual parcels. In addition, some model systems simultaneously use multiple levels of geography
that reflect transit, roadway and non-motorized impedances; however, in most cases the network
components of these model systems do not directly used this same detailed geography. The
relationships between different levels of spatial resolution in the model system should be
explicit.
Finally, most of the following discussion assumes the use of an activity-based model system
integrated with a traditional static network assignment model that converges to a static user
equilibrium solution. Other assignment methodologies may also be used in connection with
activity-based models, and the final section of this webinar considers current research linking
activity-based model systems with dynamic traffic assignment models.
966
Page 44
Representative
Population
External Trips
by Purpose
SCHOOL
(All Students)
WORK
(Student Workers)
AUTO OWNERSHIP
(Household)
SHORT-TERM CHOICE
(once per person-day)
DAY PATTERN
(activities & homebased tours for each
person-day)
Aggr.
Logsums
TOURS
(once per
person-tour)
Logsums
PRIMARY ACTIVITY
DESTINATION
HALF-TOURS
(twice per person-tour)
Aggr.
Logsums
PRIMARY ACTIVITY
SCHEDULING
ACTIVITY
LOCATION
OUTPUT FILES
MAIN MODE
PERSON FILE
(one record per
person-day
TRIP MODE
TOUR FILE
(one record per
person-tour
ACTIVITY/TRIP
SCHEDULING
TRIP FILE
(one record per
person-trip)
44
This slide shows the modules and feedback pathways in an activity-based model system (Day
Pattern Style)
967
Page 45
45
This slide shows the modules and feedback pathways in an activity-based model system (CTRAMP Style)
968
Page 46
Additional Considerations
Integration of activity-based models with other model
components
Freight/truck models
External models
Special market models (airport models/special events)
46
The personal travel demand that is represented in the activity-based model system typically
accounts for approximately 80% of roadway volumes, with about 20% of volumes associated
with freight and commercial vehicle travel, external travel with either one or both trip ends
outside the study area, or special generators such as airports. Therefore, in addition to
considering the integration of the activity-based demand components with network supply
components, it is also necessary to consider the integration of the network supply components
with these auxiliary models. Depending on the nature of the auxiliary models used to generate
this demand, the integration may be relatively straightforward or may be more involved.
For example, in some cases these demand estimates are fixed, are not sensitive to changes in
network performance, and are simply appended to the activity-based model trip just prior to
assignment. In these cases, it may simply be sufficient to ensure that the temporal and spatial
detail of this demand is consistent with that generated by the activity-based demand model.
However, in cases where these demand estimates are dynamically influenced by network
performance, it may be necessary to ensure that the auxiliary model is specified in such a way
969
that it will be sensitive to changes in network performance, and that the outputs can also be
integrated with the demand generated by the activity-based model component.
970
Page 47
47
All of the discussion above assumed that a two-way linkage was established between the
activity-based demand component and the network supply model. Despite the wide degree of
detail and fidelity that may characterize such linkages, these two-way linkages define these
model systems as integrated the demand patterns influence the network performance, and the
network performance influences the demand patterns.
It is possible to implement simpler one way linkages, but such cases dont truly represent
integrated models. There has been substantial research devoted to linking disaggregate activitybased models with similarly disaggregate dynamic traffic assignment (DTA) models. The
temporally and behaviorally detailed information produced by an activity-based model system
can typically easily satisfy the input requirements of DTA models in a way that more aggregate
trip-based models cannot. Oftentimes trip-based models outputs are significantly post-processed
and distorted in order to facilitate the use of trip-based model demand in the DTA. Activitybased model demand may not require the same degree of manipulation in order to provide the
inputs to DTA, but until a feedback loop is implemented in which the DTA provides estimates of
971
network performance for input to the activity-based model, an integrated model has not been
established.
972
Page 48
48
There are a set of core behavioral principles that address the necessity of network integration and
which also guide how network integration can be achieved. First, when making decisions,
travelers use information on expected times and costs. In an integrated model system, these
expected times and costs can be thought of as the information in the network skims that are input
to the demand model. These expected times and costs influence long-term choices such as usual
workplace school locations, as well as short-term choices such as the number and types of
activities to participate in and whether to coordinate with other household members, what
locations to visit, and what travel modes, times-of-day, and routes (roadway or transit) to use .
Collectively, these choices result in travelers new "experienced levels of network performance
which, when combined with prior expectations, then inform travelers new expectations about
travel times and costs. As a result of these new expectations, travelers may revise the long-term
and short-term travel choices. This evolution of these expectations and associated changes in
travel choices can be thought of as a learning process. Travelers will continue to adapt their
choices until the times and costs that they experience are consistent with the times and costs they
973
expected, at which point their choices will become stable. This overall iterative travel model
system may be considered an analog for this learning process.
974
Page 49
49
Model convergence is necessary to ensure the behavioral integrity of the model system, and to
ensure that the results will be useful in a policy context. The network performance or level-ofservice measurements used as the basis for accessibility measures and as key inputs to demand
model components must be approximately equal to the travel times and costs produced by the
final network assignment process. In a travel model system, there are at least two types of
convergence that we need to consider: network convergence and system convergence. When we
talk about convergence, we are implicitly talking about convergence to something. Typically
this means for networks that we are converging to an equilibrium condition (usually a
deterministic user equilibrium where, for each time period-origin-destination combination all
used routes have equal travel times, and no unused route has a lower travel time). For the overall
model system, this usually means that we are converging to a stable solution (rather than an
optimal solution as in the network context). It should be noted that in the context of an integrated
demand and network simulation model system, an essential precondition for pursuing overall
model system convergence is establishing network assignment convergence.
975
Page 50
Convergence Challenges
Mathematically, is it always possible to converge to a single,
stable solution?
Requires existence of stationary points
Uniquenessa single solution
Readily derived for aggregate demand systems using static network assignment
Theoretically possible, more difficult to achieve with microsimulation ABM-DTA
Activity-based models:
No closed analytical solution has been developed
Variations include satisfaction of agent-specific objective functions rather than
system-level objective functions-- e.g., MATSIM re-planning paradigm
Some new approaches and paradigms emerge (SHRP 2 L04)
Activity-Based Modeling: Network Integration
50
Prior to discussing different methods and measures for pursuing convergence, it is necessary to
identify some of the challenges to achieving convergence in the context of an integrated activitybased demand and network supply model system.
The first challenge is that it is not possible to confidently assert the existence of a single solution
in the context of such an integrated mode. Although it may be theoretically possible to achieve
such a solution when using a stochastic activity-based demand model in conjunction with static
network assignment model, it is not as easily provable as with a deterministic trip-based demand
model system.
While analytically derived combined demand and supply models have been formulated in
research settings, such combined models necessitate the use of many simplifying assumptions,
which compromises their usefulness in practice. In addition, even these combined models rely on
complex mathematical programming formulations.
976
In the context of activity-based models systems, no tractable analytical solution shave been
developed. Instead, both researchers and practitioners have primarily relied on heuristics such as
the method of successive averages in order to achieve stable results, although some agent-based
satisficing objective functions rather than system-level objective functions have been
successfully implemented.
977
Page 51
Achieving Convergence
Methods for overcoming stochastic variation in achieving
system convergence
Warm start the initial network assignment with trip
tables/congested travel times from a previous run
Averaging trip tables and link volumes and produced from lists of
successive iterations (similar to trip-based models)
Random number seeds enforcements
Fixing seeds for certain modules or sequences of processes
Storing arrays of random numbers generated for each module/process
51
Warm starting network assignments with trip tables of congested travel times from a
prior run;
Averaging trip tables, link volumes, or skims from successive iterations;
Fixing random number seeds and sequences that are used in the implementation of the
Monte Carlo simulation;
Using more intelligent bucket round methods such as discretizing probabilities;
Simulation or re-simulation of subsamples of the population; and
Holding some model components fixed.
978
979
Page 52
52
The method of successive averages (MSA) is a basic and reliable method for achieving
convergence in travel demand model speed-feedback procedures. The methods used in activitybased (micro-simulation) models operate on the same general principle as those used in activitybased models. The demand model produces trip tables by mode and time of day, which are run
through a static network assignment process. Common practice is then to compare changes in
link volumes or travel times, or OD interchange values between iterations.
980
Page 53
Stopping Criteria
Based on either changes to link volumes and/or trip
tables usually both
Comparisons made for each relevant demand segment
E.g., SOV, HOV, Large trucks, Small/Medium trucks
53
981
Page 54
Warm start:
3 iterations (1, , 1/3)
Input skims for Base of final (last iteration) are used as starting skims for
Build transit and highway projects
Run for Build scenarios
Hot start:
FTA New Start Methods
1 iteration only
Activity-Based Modeling: Network Integration
54
There are a number of strategies that can be used to guide the exchange of information between
the demand and supply components, in order to pursue system convergence. The selection of a
particular approach can be informed by the model application context, model runtimes, and other
factors. For example, a cold start configuration involves the assumption of some reasonable
starting skims assuming a given year and level of demand. Base case alternatives are usually
run using a cold start approach, and this strategy typically requires more demand-supply model
system iterations. Alternatively, a warm start approach uses the final base case skims to seed
the model system, which typically requires running fewer demand-supply model system
iterations. Warm starts may be more appropriate for testing build alternatives. Finally, hot
starts assume that the build alternatives pivot directly off of the base alternatives, with no
demand-supply iteration.
982
Page 55
Ongoing Research
Linking ABMs with DTA and schedule-based/dynamic
transit assignment
Current research efforts
55
As suggested earlier in the webinar, challenges arise when attempting to link disaggregate
activity based travel demand models which incorporate more detailed temporal, behavioral, and
potentially spatial resolutions with the static assignment models traditionally used in trip-based
models. A critical advantage of activity-based models is their improved temporal sensitivity, and
to fully exploit this sensitivity it is necessary to have a network supply component that can
produce realistic estimates of changes in network performance by detailed time of day. Dynamic
traffic assignment and dynamic transit assignment tools can provide this additional temporal
detail, and have many other advantages as well such as:
983
A number of research efforts are currently underway that exploring different aspects of this
problem. The SHRP2 C10 projects in Sacramento and Jacksonville are linking the Daysim
activity based model to dynamic traffic and transit assignment software and subjecting these
model systems to sensitivity testing, while the SimTravel project is developing integrated models
of the entire urban system including location choices of households and firms, activity and travel
patterns of passengers and freight, and emergent traffic flows on time-dependent networks. In
San Francisco, the SFCTA is developed a fully detailed integrated activity-based and DTA
traffic model to evaluate a variety of real-world projects.
984
Page 56
Deployment scales
Project/corridor-level
Regional-level (supports integration with regional demand
model)
Activity-Based Modeling: Network Integration
56
Dynamic traffic and transit assignment models are intended to provide more detailed means to
represent the interaction between travel choices, traffic flows, and time and cost measures in a
temporally consistent and behaviorally sound framework. More traditional static assignments
have a number of limitations, most critically their reliance of volume delay functions (VDFs).
Volume delay functions may be unreliable predictors of network performance because of their
dependence on volume/capacity ratios, which may counter-intuitively exceed 1.0, do not
incorporate critical operational attributes, do not reflect actual traffic phenomena such as
queuing, and do not reflect changes in impedances over time
In contrast, dynamic traffic assignment methods seek to explicitly represent the time-varying
nature of travel demand. This is achieved by implementing a number of fundamental departures
from static assignment methods, such as the incorporation of operational attributes which change
over time (for example, signal coordination) as well as the explicit representation of detail such
as the lane configurations. In addition, DTA models rely on traffic flow theory such as car
following and queuing to estimate delays, rather than using volume delay functions based on
985
assumed capacities. This detailed modeling of traffic flows can provide better estimates of the
true capacity and performance of different facilities based on realistic operational constraints.
At present, DTA methods are increasingly used to support project-level analyses. However, such
applications typically represent one way linkages between the demand model and the DTA
model, and in fact demand model outputs are often extensively post-processed prior to use by the
DTA model. The projects cited on the prior page represent some of cutting edge attempts to
deploy DTA approaches at a regional scale. Such a deployment is necessary in order to truly
provide useful information and integration with a regional activity-based model system.
986
Page 57
57
As described on the previous slides, DTA methods offer a number of compelling advantages
over traditional static assignment methodologies, such as more realistic representations of true
capacity constraints, operational attributes, and driver behavior. However, integration of DTA
models with regional activity-based demand models in order to achieve a fully disaggregate and
detailed travel demand models system is fraught with a number of challenges. Perhaps the most
significant challenge is that model system convergence is predicated on network supply model
convergence, but that network convergence using DTA methods is challenging. Although
analytic solutions are possible, they are extremely difficult to realize given the high computation
costs and runtimes. In practice, heuristic methods have proven to be more practical, but have not
been widely embraced by either the practitioner or research communities. And even the most
effective heuristic methods are unable to achieve levels of convergence that are easily achieved
using static methods. For example, convergence to a 2%-5% gap is typically considered good by
the DTA community, while convergence of 0.01% is usually easily achieved using static
methods.
987
Page 58
ABM/DTA Integration
Activity-based model provides demand to dynamic
traffic assignment
Disaggregate trip list
Aggregate matrices at fine temporal resolution
58
Many of the integrated activity-based and DTA model systems mirror the structure of more
traditional integrated models. That is, the demand model provides estimates of the type, timing,
destination, and model used for travel, while the supply model takes these estimates of demand
as fixed and uses them to generate estimates of network performance. Of course, the fully
disaggregate nature of both tools supports the implementation of unique model integration
methods. The activity-based demand model may feed the DTA model a trip list, or may
aggregate this trip list to spatially and temporally fine grained matrices. The DTA model, in turn,
can provide more detailed and consistent estimates of network performance by time of day. At
present, these tools have achieved integration using skims as detailed as 48 hour time periods,
although in theory skims may one day be generated on the fly rather than developing a full set
of segmented skims prior to the demand model simulation.
988
Page 59
LOS skims
for all
possible
trips
Trip tables
Static assignment
59
This is a reminder of the essence of conventional integration scheme in a 4-step model. Note that
trip tables and LOS skims are in exactly the same matrix format.
989
Page 60
Microsimulation ABM
Individual
trajectories
for the
current list of
trips
List of
individual
trips
LOS for
the other
potential
trips?
Microsimulation DTA
60
The same idea does not work for DTA-ABM integration directly. The crux of the problem is that
the feedback from DTA to ABM is not straightforward. DTA only produces individual
trajectories for the current set of trips. This is not enough to support an ABM across all travel
choice dimensions and for all potential trips.
990
Page 61
Possible Surrogate
Microsimulation ABM
Aggregate
LOS skims
for all
possible trips
List of
individual
trips
Microsimulation DTA
61
A possible surrogate solution that solves some problems in practical terms is to use aggregate
LOS skims instead of individual trajectories. This so far has been the main strategy, but it must
be understood that a lot of individual information is lost in the aggregation of these data.
991
Page 62
Individual
trajectories
for potential
trips
Consolidation of individual
schedules (inner loop for
departure / arrival time
corrections)
Individual
trajectories
for the
current list of
trips
List of
individual
trips
Microsimulation DTA
Temporal
equilibrium
to achieve
individual
schedule
consistency
62
New methods of equilibration for ABM and DTA are presented, where two innovative technical
solutions are applied in parallel. The first solution is based on the fact that a direct integration at
the disaggregate level is possible along the temporal dimension if the other dimensions (number
of trips, order of trips, and trip destinations) are fixed for each individual. Then, full advantage
can be taken of the individual schedule constraints and corresponding effects. The inner loop of
temporal equilibrium includes schedule adjustments in individual daily activity patterns as a
result of congested travel times being different from the planned travel times. It is very much
helps the DTA to reach convergence (internal loop), and is nested within the global system loop
(when the entire ABM is rerun and demand is regenerated).
The second solution is based on the fact that trip origins, destinations, and departure times can be
pre-sampled and the DTA process would only be required to produce trajectories for a subset of
origins, destinations, and departure times. In this case, the schedule consolidation is implemented
though corrections of the departure and arrival times (based on the individually simulated travel
times) and is employed as an inner loop. The outer loop includes a full regeneration of daily
992
activity patterns and schedules but with a sub-sample of locations for which trajectories are
available.
993
Page 63
Ti
Duration
di
Arrival
i
i
Departure
Schedule
q i
Activity i=0
Activity i=1
Trip i=1
Activity i=2
Trip i=2
Activity i=3
Trip i=3
24
63
Individual schedule consistency means that for each person, the daily schedule (i.e., a sequence
of trips and activities) is formed without gaps or overlaps as shown on the slide. In this way, any
change in travel time would affect activity durations and vice versa.
994
Page 64
x
y
min xi ln i yi ln i
di
i
i
Previous
durations
Daily consistency
New
departures
Previous
departures
x t 24
i
Departure time
yi x j t j
j i
Solution
Activity-Based Modeling: Network Integration
xi k di
j 1
Changed
travel times
j
yj
64
995
Page 65
65
This model was further extended to incorporate various schedule constraints and priorities. The
essence of this formulation is that in presence of travel times that are different from the expected
travel times that the user used to build the schedule, he will try to accommodate new travel times
in such a way that the schedule is preserved to the extent possible. The preservation relates to
activity start times (trip arrival times), activity end times (trip departure times), and activity
durations. The relative weights relate to the priorities of different activities in terms of start time,
end time, and duration. The greater is the weight, the more important for user to keep the
corresponding component close to the original schedule. Very large weights correspond to
inflexible, fixed-time activities. The weights directly relate to the schedule delay penalties as
described below in the section on travel time reliability measures. However, the concept of
schedule delay penalties relates to a deviation from the (preferred or planned) activity start time
(trip arrival time) only, while the schedule adjustment formulation allows for a joint treatment of
deviations from the planned start times, end times, and durations.
996
Page 66
Duration
5
5
20
1
1
1
1
5
1
5
3
1
1
Trip departure
(to activity)
1
1
1
1
5
10
5
1
1
5
3
5
1
66
This is an example of typical weights applied in the schedule consolidation algorithm. These
weights can be further individualized. Note the extreme cases with high weights that correspond
to fixed schedule components. Lower weights correspond to flexible schedule components.
997
Page 67
67
This method is intended to resolve one of the fundamental problems associated with integration
of micro-simulation ABM and DTA the calculation of individual LOS variables for nonobserved destinations and times of days (i.e. for trips that were not simulated at the previous
global iterations). It is also behaviorally more appealing to assume that an individual does not
always scan all possible location in the region for each activity but rather operate within a certain
spatial domain where he explores options over time and make choice based on the past
experience.
998
Page 68
68
Pre-sampling of destination constrains the variation of destinations for each individual and
allows for an efficient accumulation of individual trajectories in the micro-simulation process.
With this technique, the LOS variables will be defined at each subsequent iteration as follows:
First, individual trajectories to the same destination by departure time period for the same
driver (or some other driver from the same household) are used if present in the previous
simulation; behaviorally, this corresponds to personal learning experience; having only
30 possible destinations enhances this probability for each individual; if not:
Individual trajectories to the same destination by departure time period across all
individuals are used if present in the previous simulation (if several of them are available,
the average can be used); behaviorally this corresponds to social networking when the
driver can hear from other people about their experience; having only 300 possible
destinations for each origin MAZ enhances this probability, if not:
Aggregate LOS skims by departure time period will be used as the last remaining option;
behaviorally it can be thought of as using an Advice from an advanced navigation device.
999
Page 69
Auto
SOV
HOV2 driver,
joint travel party
HOV3+ driver,
joint travel party
Transit
HOV2 passenger
(not assigned)
HOV3 passenger
(not assigned)
Non-toll, General
Purpose lane
Non-toll, General
Purpose lane
Non-toll, General
Purpose lane
Non-toll, General
Purpose lane
Non-toll, General
Purpose lane
Toll, Managed
lane
Non-toll,
Managed lane
Non-toll,
Managed Lane
Non-toll,
Managed lane
Non-toll,
Managed lane
Toll, Managed
lane
Toll, Managed
lane
Toll, Managed
lane
Toll, Managed
lane
69
1000
Page 70
70
In order to make the ABM compatible with the temporal resolution of DTA, we suggest an
enhancement of the temporal resolution of the trip departure time choice model to 5 min. The
tour-level time-of-day choice model is characterized by a complicated two-dimensional choice of
outbound and inbound times that results in approximately 1,000 alternatives with 30-min
resolution. This choice dimension is difficult to further improve in terms of temporal resolution.
Also, in behavioral terms, there is no need in too much temporal detail in tour-level scheduling
anyway, since it does not yet have full information on the number/location of destinations visited
along the tour, so the exact timing and LOS information is only indicative at this point in the
simulation. However, trip-level choice of departure time that is conditional upon the entire-tour
time-of-day choice can be refined to 5 min, since the choice structure is one-dimensional and
more details about each particular trip origin and destination can be used.
1001
Page 71
71
One of the more ambitions research efforts currently underway is the FHWA sponsored
SimTravel project. This wide-ranging project is developing integrated models of the entire urban
system including location choices of households and firms, activity and travel patterns of
passengers and freight, and emergent traffic flows on time-dependent networks. Unlike
integrated models which seek to achieve an equilibrium condition by the repeated exchange of
demand and supply data, much of the early SimTravel work around demand and supply
integration is focused on real time interactions between the demand and supply components.
Ultimately, such an approach seems as though it may require fully integrated demand and supply
models in a single program structure. However, such an approach may also offer compelling
application and analysis capabilities.
1002
Page 72
72
There have been several recent initial attempts to use actual/realistic transit schedules in
conjunction with activity-based models and dynamic traffic assignment. Specific efforts have
included attempts to model the following behavior:
1003
Page 73
73
The most significant benefit of an activity-based model system that has been carefully integrated
with a network supply model is improved policy sensitivity and by extension, greater confidence
when comparing investment and policy alternatives. This improved policy sensitivity is critical,
as decision-makers rely on travel demand models for answers to more complex policy questions,
and expect that these models are appropriately sensitive to the complex behavioral responses.
Whether evaluating the impacts of a pricing scenario on the use of specific routes, timing, mode,
destination or activity generation choices, or considering the impact of land use or travel demand
management strategies, these complex questions necessitate models that appropriately and
consistently capture the relationship between travel demand and supply. If network pricing is to
change by time-of-day, then ideally network information (times, costs) at a time resolution
consistent with this pricing scenario can be fed back from supply model to demand model.
Similarly, if value-of-time distribution information is used when predicting demand, then ideally
this segmentation would be reflected in the configuration of the supply model. Greater detail and
1004
greater consistency between the demand and network supply components of the model system
lead to improved behavioral realism. In addition, this greater detail in both the demand and
supply components supports an expanded set of performance measures.
1005
Page 74
Runtime:
Stochastic variation may necessitate multiple demand
simulations
When ABM is integrated with static UE assignment, more
detail/resolution means longer runtimes
More time periods to assign
More market segments to assign
74
However, this improved policy sensitivity comes at a cost. From a data management perspective,
the development and maintenance of network inputs does require additional effort, and more
fine-grained calibration and validation efforts are also associated with this additional detail.
From a production standpoint, the additional time period and market segmentation results in
longer runtimes when using static assignment methods, although these runtimes differences
become less significant with improvements in network software and increases in hardware
computing power. Within the context of DTA, these runtime differences associated with
additional may detail become even less relevant. Depending on the specific application context,
stochastic variation associated with the use of the Monte Carlo simulation based activity model
may also necessitate multiple simulations and additional runtime, although the demand
components of most activity-based model systems typically account for only a small fraction of
total model runtimes.
1006
Page 75
1007
Page 76
Next Webinar
Executive and Management Sessions
Executive Perspective
Institutional Topics for Managers
Technical Issues for Managers
Technical Sessions
Activity-Based Model Framework
Population Synthesis and Household Evolution
Accessibility and Treatment of Space
Long-Term and Medium Term Mobility Models
Activity Pattern Generation
Scheduling and Time of Day Choice
Tour and Trip Mode, Intermediate Stop Location
Network Integration
Forecasting, Performance Measures and Software
February 2
February 23
March 15
April 5
April 26
May 17
June 7
June 28
July 19
August 9
August 30
September 20
76
1008
the way the models are interacted with each other. So, while you do expect the convergence to
go hand in hand, there are instances where it cant be assumed, depending on the nature of the
implementation of the model.
This question concerns the potential of using DTA instead of static assignment, and what we do
if we are not using volume-to-capacity (V/C) ratios. Could V/C ratios be capped during static
assignment to better represent capacity constraints? If we are doing away with V/C ratios, what
criteria do we use to establish level of service?
Peter: This goes back to better understanding the differences between static assignment and
volume delay functions versus the simulation techniques such as car-following and queuing rules
in a dynamic assignment. There have been many publications showing that V/C ratios cannot
reliably establish travel time estimates, especially when V/C goes over one. This mechanism is
replaced by simulation in a DTA, and when volume approaches capacity, a variety of things
happen such as gridlock and queue spillback that more realistically address reality and do not
allow V/C to rise above one. The publications show that the DTA methods more accurately
estimate travel times during congestion. Whenever you try to just cap V/C to get more accurate
travel times or stronger volume-delay functions, which we have tried to do, it doesnt work. The
right way to go is have a tool such as DTA that actually represents the physical situation on the
road. We need a new approach, but I do understand the concern for simplified practical solutions.
While I dont recommend capping V/C, you can scale time estimates from static assignment
based on observed values to provide a better matrix. But this is a temporary, surrogate solution,
and I think we need to take DTA seriously. This is the way to go.
1010
1011
Page 1
Activity-Based Modeling
Session 12: Forecasting and Application
1012
Page 2
Acknowledgments
This presentation was prepared through the collaborative efforts of Resource
Systems Group, Inc. and Parsons Brinckerhoff.
Presenters
John Gliebe & Peter Vovsha
Moderator
Stephen Lawe
Media Production
Bhargava Sana
This series is and has been a collective effort between RSG & PB. It is largely built on our
experience with many activity-based models in practice.
Your presenters for today are John Gliebe and Peter Vovsha. Maren Outwater actually prepared
most of the content for this presentation, along with Peter and John. The slides at the end of the
presentation showing the visualizations from the Atlanta Regional Commission activity-based
model were developed by Joel Freedman and Ben Stabler (PB). We were also supported by Mark
Bradley, who reviewed the presentation. The multi-media production for this webinar is being
handled by Bhargava Sana.
1013
Page 3
February 2
February 23
March 15
April 5
April 26
May 17
June 7
June 28
July 19
August 9
August 30
September 20
3
Today webinar will cover Forecasting, Performance Measures and Software. It is a natural
follow up to the previous webinar on activity patterns. This is the final webinar in the series. We
invite you to return to the TMIP webinar archives if youre interested in viewing and listening to
any of the previous webinars in the series.
1014
Page 4
Learning Outcomes
Steps involved in preparing an activity-based modeling system
for forecasting
Output measures produced with an activity-based modeling
system
Differences in alternatives analysis with an activity-based
modeling system
Hardware and software considerations for activity-based
modeling applications
At the end of this session, participants should be able to answer the following questions:
What are the steps involved in preparing an activity-based modeling system for
forecasting?
What output measures are produced with an activity-based modeling system?
What are the differences in alternatives analysis with an activity-based modeling system?
What are the hardware and software considerations for activity-based modeling
applications?
1015
Page 5
Downward
Integrity:
Choices
made in
higher
models affect
choices made
in lower
models
Mobility
Choices
Model
Inputs
Long-Term
Choices
Daily Activity
Patterns
Tour & Trip
Details
Trip
Assignment
Model
Outputs
Upward
Integrity:
Expected
utility of
making
choices in
lower models
affect choices
made in
higher
models
Before going further, it is worth revisiting where weve been. This being the final webinar in the
series, we have now covered all of the fundamental components of activity-based modeling
system. This includes the creation of a synthetic populations and important model inputs, such as
the treatment of space and accessibility; long-term and mobility choices; activity pattern
generation; tour and trip details such as scheduling and time-of-day choices, and tour and trip
destination and mode choices; and finally network assignment and its integration. Throughout
this series, weve emphasized the integrity of these modules as a working whole. Downward
integrity refers to choices made in upstream models conditioning the choices made in
downstream models. Upward integrity is where the expected maximum utility of these
downstream choices affects the probabilities of the choices made first, in the upstream models.
Now we are ready to discuss model outputs in more detail.
1016
Page 6
Outline
Importance of forecasting methodology, performance
measurement and software development
Basic terminology
Calibration with activity-based models
Performance measures and sensitivity testing
Random variation and alternatives analysis
Example applications
Implementation in hardware and software
Areas of research
6
In this session we will discuss various aspects of the forecasting process, highlighting the
differences between activity-based models and trip-based models. We will begin by discussing
the importance of methodology, performance measurement and software development. We will
then define some basic terms with which you should be familiar to better understand forecasting
with activity-based models. Next, we will talk about calibrating an activity-based model and the
additional considerations that that entails. After that we will discuss performance measures and
sensitivity testing and why this is where activity-based models reveal their true explanatory
power. We will then discuss how to handle random variation across model runs, particularly as it
pertains to alternatives analysis. We will review some example applications of activity-based
model forecasts. We will discuss important hardware and software issues. Finally, we will spend
a little bit of time discussing areas of research.
1017
Page 7
There are several themes that wed like to emphasize today in order to convey the important
differences between activity-based model and trip-based model forecasting practice. First, with
activity-based models, a much richer array of output measures are possible, which you shall see
later when we discuss performance measures. Second, the internal complexity requires new
understanding of how to properly calibrate and validate activity-based models. This is because
there are more model components and they have a larger number of interdependencies. In
addition, activity-based models are not just about tripsinterpretation of daily patterns, tours,
activity durations are important to comprehensible forecasts. The use of simulation to produce
forecasts makes it necessary to control and explaining random variation, which is important to
producing consistent forecasts and communicating with decision makers. Finally, the added
complexity of activity-based models means that new application software is needed, both to take
advantage of more powerful analytical capabilities and to handle greater computational loads.
1018
Page 8
Terminology
Micro-simulation
Performance measures
Forecasting
Data visualization
Multi-threading
Distributed processing
This slide defines some of the key terms as they are used in this presentation.
Micro-simulation in this sense refers to a travel demand model that simulates individual agents
(person, households, and vehicles).
Performance measures are outputs of the travel demand model that may be used to assess the
benefits of a strategy or alternative.
Forecasting refers to representation of a future year with assumptions about growth,
transportation and the economy.
Data visualization includes graphic, tabular or spatial presentation of model output or input.
Multi-threading is processing across multiple cores within a computer.
Distributed processing refers to processing across several computers in a network.
1019
Page 9
Steps Involved
Preparing forecasts involves steps very similar to those
of trip-based model development
Base-year calibration and validation
Horizon-year baseline forecasts and sensitivity tests
Alternative forecasts
Network alternatives
Land use alternatives
Policy alternatives
Preparing forecasts involves steps very similar to those of trip-based model development. We
always start with the development of a calibrated and validated base-year model. As we are
about to discuss, there are some similarities to trip-based model calibration and validation, but
some important differences as well. From there, it is typical to create horizon-year baseline
forecasts. When an agency first develops an activity-based model, it is common to compare the
forecasting results to those of their trip-based model. This helps the agency to spot differences
and to explain those differences. Some general sensitivity testing is also recommendedvarying
certain input assumptions, for exampleto make sure that the model system responds as
expected. Once an activity-based model is put into operation, there are potentially many different
types of uses to which it may be applied. Starting out, an agency may want to try it out with
some basic types of alternative analyses such as: network alternatives, land use alternatives, and
various policy alternatives. It is important for an agency adopting a new tool to become
comfortable with the way it behaves, to do any necessary fine tuning, and to know what to
expect when it comes time for the next big long-range plan update, conformity analysis, or
policy study.
1020
1021
Page 10
Base-year calibration and validation may be described by two basic phases. In the first phase,
individual model components are calibrated while holding others constant. Typically, you would
be estimating model components from household activity-travel diary data. The survey would be
expanded and calibration target values set for all of the model system components. In addition,
you may want to supplement the survey data with other sources, such as the Censuss journey to
work data, which should be a more comprehensive indicator of commuting trip lengths and
spatial patterns. Another consideration is the activity-based model systems will have many more
component modules than a 4-step model, so there will be more points of calibration. While this
adds to the amount of time required for calibrating the model, it allows the model system to
respond in more behaviorally consistent ways when applied to various scenarios that happen to
deviate from the baseline.
The next phase of calibration is system-level with feedback. Here we want to makes sure that the
individual model components remain in calibration, although there will usually be additional
adjustments made. In general, we do not expect an activity-based model to do a better job of
matching traffic counts or boarding counts than a trip-based model. If two models started from
1022
scratch, with no calibration, the activity-based model generally gives a more reasonable first pass
results. But, usually the comparison is to an existing trip-based model that has already been
(over-calibrated). It is typically easier to calibrate an activity-based model without extensive use
of K-factors.
1023
Page 11
Iterative Calibration
Model Inputs
Compare to observed
Assess Reasonableness
of Forecasts
Highway volumes
Transit boardings
Model
Outputs
Synthetic
Population
Trip
Assignment
Trips
Destinations
Modes
Long-Term
Choices
Mobility
Choices
Daily Activity
Patterns
Households
Persons
Home location
Work location
School location
Auto ownership
Transit pass
Tours
Purposes
Schedule
11
This figure illustrates the iterative calibration process that I just spoke of. It is also a good
refresher for talking about activity-based model components. The synthetic population is the first
thing to calibrate; however, this can be done essentially independently of the other model
components, and the methods used in synthetic population generation software do the calibration
for you, using marginal control totals. From there, we proceed down the model stream, first
calibrating the long-term choice models, which condition the mobility choice models. Long-term
workplace and school location choice are usually inputs to mobility models such as auto
ownership, transit pass holding, and the like. The results of all of these long-term and mobility
choices are inputs to the daily activity pattern choices and to models of tour and trip details
destinations, modes and time of day. We want to calibrate tour frequencies by purpose,
destinations, modes, and time of day choices prior to calibrating trip-level characteristics. Trip
frequencies, destinations and modes are typically the last models in the sequence to be calibrated.
Once we have gone through this sequence, we would then apply the entire model system and
obtain assignment results. If we are at the stage of validation, wed be making comparisons to
traffic volumes and transit boardings. In addition, we may make some additional checks on the
1024
reasonableness of the activity-travel patterns being forecast. Well discuss more of about
reasonableness checks later in the presentation. Having gone through the process once, it is likely
that some of the model components that we calibrated first will no longer be in calibration if we
feed back the assignment skims. We will step through this process for at least one other iteration,
and possibly multiple additional iterations, until were satisfied that most of the calibration
targets have been met and are stable.
1025
Page 12
12
Some calibration metrics are familiar to trip-based modelers and can be derived from activitybased model output. These include:
1026
Page 13
13
In addition, however, tour or activity-based models require that an additional set of model
predictions be calibrated against benchmark values. Typically, these include:
1027
Page 14
While activity-based models may take more time to calibrate because they have many more
model components than trip-based models, practice has shown that they tend to result in fewer
problems both in validation and in other types of scenario/sensitivity runs. This because you
have more calibrated support points for forecasts. This is partially due to modeling a
disaggregate population and partially due to all of the linkages between model components. So,
investment in finer resolution and linked decisions pays off by putting the calibrated model in a
better starting position.
1028
Page 15
Transferring an activity-based model system from one region to the next has become a popular
alternative to developing model components from scratch because it can get an agency up and
running faster and presumably makes use of a model design and structure that has worked
elsewhere. The jury is still out as to whether this is a good idea. The first successful example
of this is probably the transfer of the Columbus, Ohio (MORPC) activity-based model to the
Lake Tahoe region. This is a somewhat unusual case, given the characteristics of the Tahoe
region as a resort area. The Atlanta Regional Commission and Metropolitan Transportation
Commission (S.F.-Oakland Bay Area) is an example in which the CT-RAMP model structure
was developed jointly for both regions, although it appears the model coefficients were not so
readily transferred. As another example, the Sacramento model was transferred to the Fresno and
North San Joaquin Valley region, which makes more sense because they are adjacent regions.
Commonly, the idea is to transfer an existing model to get it running and start to get the staff
familiar with it, but then to recalibrate it using local data. In the recent case of the Delaware
Valley Regional Planning Commission (Philadelphia area), the model transfer from the Puget
1029
Sound Region (Seattle) has just begun, and the long-term plan is to re-estimate the model system
in its entirety once a new household survey has been completed.
FHWA is quite interested in the issue of transferability and has a commissioned research on this.
In the STEP project, comparisons are being made between several regional models of the
DaySim framework have been developed for Sacramento, Shasta County, and Fresno, California
as well as for Jacksonville and Tampa, Florida.
1030
Page 16
Intuitive interpretations
Ability to trace back outcomes to their source of change
Improved communication with planners
16
The true power of activity-based models tends to come out when doing disaggregated forecasts.
For example, a trip-based model might not account for a one-way toll applied in the PM peak
period, because it does not affect the AM Peak skims, whereas an activity-based model would
model the travel impedances in both directions and predict a more appropriate response. Another
example would be the ability to test travel demand management policies, such as shortened work
days or work weeks.
In general, activity based models, being disaggregate representations of travel behavior, give the
analyst the ability to trace travel behavior outcomes back to the source of their change. Of
course, one needs to know where to look! And thats where training and familiarity come in. In
addition, activity-based model facilitate better communication between modelers and planners
because they can point to more intuitive explanations for changes in travel behavior.
1031
Page 17
Equity
ABM output looks like a full population household survey
and can be expanded in any way to understand equity
17
Traditional trip-based transportation system performance measures are still available in activitybased models. These include things like trip length distributions, mode shares, and link-based
and OD-based travel times and costs. In addition, however, an activity-based model can do a
better job of predicting impacts on individuals within the population. For example, this is
especially useful when analyzing equity.
1032
Page 18
18
Lets look at some examples. Here is an example of total daily travel time predicted by a
Chicago regional model in which the data are summarized by household income group and by
person types. Here you see the four income group levels in the black type with the black dot
along the horizontal dashed line, and if you follow this you can see that total daily travel time
(per capita) increases as you go up the income levels. In fact there is nearly a 20-minue
difference from the lowest group to the highest group.
Similarly, this graph shows large differences in daily travel minutes for different person types.
Full- and part-time workers have the highest values, followed by university students, driving age
secondary school students, non-working adults, non-driving age students, retirees, and lastly preschool age children.
1033
1034
Page 19
19
Here is another example. Zero vehicle and low income households benefit from travel time
savings where non-zero and non-low income households do not benefit. This may be due to the
focus on transit investments and the fact the low income and zero-car households are more likely
to happen in transit rich areas. Female head of household with children and single parent
households do not have as much travel time savings as the average household. This may be
because transit investments are serving work trips in transit rich areas, where female and single
parent households are not traveling as much.
1035
Page 20
Growth
Jobs-Housing Balance
Growth in centers
20
Activity-based modeling can also help us evaluate environmental performance measures. This is
because disaggregate modeling allows us to trace travel back to the household and individual
sources and because it provides more precise estimates of travel demand by time of day. Because
we are also modeling activity duration, we have better estimates of hot and cold starts. If an
activity-based model is integrated with a dynamic traffic assignment (DTA) program, you can
even get more precise estimates by facility type and time of day and trace vehicles. In fact, future
activity-based models are expected to model vehicle usage by engine and fuel type.
For land use and growth management scenario testing, it is useful to test the impacts of proposals
on activity-travel patterns. Activity-based models are designed to be sensitive to accessibility.
This makes them more sensitive to commuting distances and modes, as well as discretionary trip
making. For example, an activity-based model can capture the phenomenon of persons with
reduced commuting times and costs taking advantage of this time by making additional
discretionary trips.
1036
Page 21
21
This is an example and a valuation of energy use by alternative and source that was completed as
part of a long-range planning effort in Seattle. In this example, fuel and electricity use was
estimated from the vehicle miles traveled produced by the travel demand forecasting model and
from square footage estimates of buildings from the land use forecasting model. This comparison
shows significant reductions in fuel consumption for several planning alternatives and very little
change in electricity use from the buildings. This evaluation was designed to also account for
electricity use from electric vehicles, but was not employed for these planning alternatives.
1037
Page 22
22
Another example shows vehicle miles traveled from the Sacramento activity based model on a
per household basis. Tracking vehicle miles traveled on a per household basis is an advantage
over trip-based models which can only track VMT on a link basis. This allows us to understand
the source of vehicle miles traveled which is important because policies would affect households
rather than links. In this example, the VMT with density at the household level and growth in
areas of higher density will have lower VMT per household. Models that track VMT by link will
show that VMT is higher in areas of higher density which is why it is so important to track VMT
at the household level.
1038
Page 23
Quality of Life
Safety (crashes)
Health (active transportation)
23
Another category of performance metrics describes economic development and quality of life.
One common measure is the user benefits accruing to various economic sectors. One example of
user benefits for economic development include user benefits to low- or high-wage employment
sectors, which measures equity and potential for prosperity respectively. Another example are
user benefits to freight centers. This measures freight mobility, which in turn is an indicator of
the potential to attract more freight to these centers.
There are several ways to measure quality of life emanating from the transportation system. A
measure of safety is the number and severity of crashes occurring on the roadway system. A
measure of health is the number of non-motorized trips as an indicator of the number of people
engaging in active transportation modes.
1039
Page 24
24
This graph shows these same economic prosperity measures and the extent to which different
economic development plans produce user benefits. Cluster employment refers to those sectors
that the regional economic development plan has identified as showing promise for economic
prosperity. Freight employment covers those sectors that serve freight mobility and are
therefore preferred for positive user benefits. High-wage employment benefits will attract more
high-wage employment to the region. As you can see, there are seven alternatives and the last
two which are the light and dark gray bars, respectively, would seem to confer the most benefits
and are therefore the preferred alternatives.
1040
Page 25
25
Activity based models have several advantages for non-motorized travel over trip based models
due to the additional spatial detail and the disaggregate nature of the models. Some
considerations for modeling bike and walk trips that are on both the demand and supply parts of
the modeling system. These include representing bicycle and pedestrian facilities in the
networks, separating bike access to transit from walk access to transit, including bike and walk
times in accessibility measures along with auto and transit times, representing walk and bike
routes on the network from parcel to parcel instead of a more simplified straight-line distance
estimate, including bike route choice model components, and using distance decay functions in
buffer accessibility measures rather than a simple half-mile threshold. Any or all of these
considerations can be included in the activity based modeling system to improve the accuracy of
walk and bike trips.
1041
Page 26
Scenario Testing
Reasonableness tests -- sensitivity of outputs to
changes in inputs
Evaluation of specific policies and projects -- is the tool
appropriately specified for the job for which it is
needed?
Fine tuning assumptions and specifications
26
Scenario testing is an important means to assess the reasonableness and appropriate use of
activity based models. Scenario testing can include testing the sensitivity of outputs to various
changes in the inputs as well as the evaluation of specific policies and projects. This can be
useful to understand whether the activity based model has been appropriately specified to meet
the needs of any specific planning application. Scenario testing can also be used to fine-tune the
assumptions and specifications within the models. Sensitivity testing is often used to assess the
reasonableness of specific policies before the model is used to evaluate the effect of a specific
project which may be more difficult to interpret due to the complexity of the project details.
1042
Page 27
27
Scenario testing was used extensively to evaluate parking pricing scenarios in San Francisco.
This example shows the results of a three dollar charge into and out of a Northeast cordon,
shown here in blue, during weekday peak periods; compared with a three dollar parking charge
in a focused area, shown here in green. The Northeast cordon charge reduced vehicle miles
traveled citywide more than the focused area parking charge due to the broader coverage of daily
trips that were charged. This scenario also increased walk and bike trips much more than the
focused area scenario, but did not increase peak period transit trips as much. These parking
pricing scenarios would be quite difficult to model using a trip-based model.
1043
Page 28
Sensitivity Testing
Sensitivity of outputs to changes in inputsbaseline vs.
alternative scenario
Reasonableness of shifts in
Routes, time of day, destinations, modes
Consider place-to-place geography, neighborhood impacts
Consider differences by value of time segmentation (activity purpose,
income)
Consider differences by vehicle availability segmentation
Are there certain daily patterns that are under- or overpredicted (relative to survey)?
28
Sensitivity testing is the process of changing model inputs to evaluate the effect on outputs.
Typically, we change model inputs using simple factors so that the outputs can be interpreted
more easily. For example, a change in a bridge toll could be tested by doubling the current price
rather than implementing a more specific evaluation of a dynamic or time of day cost structure
on the bridge. Sensitivity testing can be used to assess the reasonableness of shifts in routes, time
of day, destinations, modes, or locations. We can also summarize the reasonableness of model
outputs with respect to value of time segments or vehicle availability segments of the population.
Reasonableness testing is as much art as science. Basically, one needs to consider the magnitude
of the shifts produced by the modelshifts in route, time of day, O-D patterns, and mode
and assess whether they make sense and are similar to what has been observed in the past when
such changes occur. For new policiessuch as pricingit would be ideal to look at places
where pricing changes have occurred to get some idea of what to expect. In most cases it comes
down to professional judgment and requires a high level of experience with modeling in
generalnot necessarily activity-based models.
1044
Page 29
29
This example from the Oregon statewide model shows a sensitivity test on driving cost which is
four times the original input shown in red and 10 times the original input shown in blue. The
results shown here demonstrate that regional center densities will increase 20% under the higher
cost options and other areas will have density gains up to 10%. Are these results realistic? For
example, if gasoline prices increased to $15 per gallon or $38 per gallon is this the change in
land use density that wed expect to see? How long would the land market take to adjust? These
are difficult questions to answer when we have not experienced such costs.
1045
Page 30
One issue in travel demand forecasting that can be better assessed in activity based models are
the stability across scenarios. For example research shows that people maintained time budgets
in daily life so that total amount of time spent on travel is about the same and increases in one
type of travel will result in decreases in another type of travel. It is useful then to review
forecasts of average activity durations to determine if they are stable and reasonable over time.
The details of various scenarios like vehicle hours of travel or walk access and egress distances
to transit can provide insights as to the stability across scenarios.
1046
Page 31
31
Travel demand forecasting models should be dynamic in the sense that they will be improved,
assessed, and updated over time. In practice, travel demand forecasting models will typically be
updated for a specific purpose which improves all aspects of the system needed for that purpose,
but it may not cover all other purposes until there is a need. It is important to assess the
reasonableness of the model specifications for a particular analysis. For example, evaluating road
pricing options requires that the model responds appropriately to price signals and produce
elasticities for route choice, time of day, intermediate stops, modes, and work location choice
that are consistent with our expectations.
1047
Page 32
32
For each planning application, it is useful to fine tune assumptions and specifications in the
activity based model as a result of sensitivity tests specifically aimed at understanding the model
in the context of the specific planning application. Counter intuitive results or non-response
should lead to changes in the assumptions or specifications of the model to correct these results.
This can lead to changes in input data or assumptions or specifications for a specific choice
model.
It is also useful to consider the confidence we have in forecast variables and whether we need
these forecast variables for policy testing. Sensitivity testing can be used to including variables
for policy testing that we may have less confidence in forecasting.
1048
Page 33
33
Activity-based models have some clear advantages over trip based models because of their
disaggregate data throughout the system. This allows us to summarize any household or person
attribute along any geographic stratification or time of day. Trip based models are limited to
summarizing outputs by the segments that are directly represented in the model. Simulating
individual people has clear theoretical advantages but also presents a practical challenge in
handling the stochastic effects and producing a slightly different outcome each time the model is
run. Many have found that an alternative analysis is easier to understand and interpret if these
stochastic effects are eliminated, but some people feel that the range of results from these
stochastic effects is useful to understand.
1049
Page 34
Stochastic Variation
Advantage: ability to portray a distribution of outcomes
more realistic ability to portray risk
Challenges:
Need to demonstrate that random variation does not swamp
meaningful changes in policy variables
Non-technical decision makers may prefer a single number
Some analyses require analysis of comparative statics
34
Stochastic variation has a clear advantage in portraying a distribution of outcomes and the
uncertainty associated with these outcomes, which is more realistic than a single number. The
challenge, however, is to demonstrate that this random variation does not overshadow
meaningful changes in policy variables. Another challenge is that many decision-makers find
this confusing and prefer a single number result that can be more easily interpreted and
compared across planning alternatives. Note that methods have been developed, like random
sequence synchronization, to minimize the random stochastic differences BETWEEN scenarios,
which addressed the first challenge.
1050
Page 35
Number of Iterations
Consider changes across
multiple outcome variables
Link volumes
Mode shares
Tour lengths
Trip lengths
Work destinations
Other destinations
etc.
35
1051
Page 36
36
This example from San Francisco shows the random simulation error that occurred over 100 runs
of the activity based model. The graphic shows that aggregated results at the county or
neighborhood level have much less random simulation error then results at the traffic analysis
zones level.
1052
Page 37
37
There are several strategies for controlling stochastic variation that can be deployed. This
includes constraining random number sequences by starting from the same random number seed
or by saving random number sequences to eliminate the random numbers used in the activity
based model. Another strategy is to run the model multiple times and average the trip tables from
these runs for use in assignment. A third strategy is to hold certain model components fixed
between model runs. There are also feedback and convergence issues with integration of network
models that was discussed in the previous webinar that should be considered.
1053
Page 38
38
1054
Page 39
New Starts
Pricing studies
Conformity
Regional Transportation Plans
Environmental Impact Statements
39
The next part of the webinar covers the types of activity based model applications with several
examples, software platforms, and integration topics for model applications. There are many
types of activity based model applications. I will cover four of the most common and show
several examples. These include transit ridership forecasting for new starts applications,
conformity analysis, forecasting of alternatives for regional transportation plans and the
evaluation of forecasts for environmental impact statements.
1055
Page 40
MORPC AB model:
COTA North Corridor LRT/BRT Study
NYMTC AB model:
Tappan Zee Bridge Study
40
Current news starts modeling requires that the number of trips and their locations be fixed across
alternatives and that highway skim values remain the same across alternatives. This requires
adapting the activity based model so that outputs from the baseline model run can be used across
all build alternatives allowing us to isolate mode choice differences. New starts applications
should reuse random numbers generated in the baseline case for all build alternatives so that
change is only attributable to the differences in utilities.
1056
Page 41
Major Issues
FTA fixed trip table requirement & mode choice
logsum as UB measure:
What to fix in tour-based structure?
Where to calculate UB in the model chain?
41
Applying activity based models for new starts applications involved fixing trip tables to ensure
that mode choice and assignment results could be compared more directly across alternatives.
New starts applications for activity-based models have explored interpreting tour and trip level
outcomes in mode choice. Mode and destination choice at the tour and trip level need to be
accounted for without double counting user benefits. This requires consistency in defining
markets at the tour and trip levels. The simulation of individuals provides flexibility in defining
market segments.
Current news starts modeling requires that the number of trips and their locations be fixed across
alternatives and that highway skim values remain the same across alternatives. This requires
adapting the activity based model so that outputs from the baseline model run can be used across
all build alternatives allowing us to isolate mode choice differences. New starts applications
should reuse random numbers generated in the baseline case for all build alternatives so that
change is only attributable to the differences in utilities.
1057
Page 42
Fixed across
all scenarios
Trip distribution
Time of day
Mode choice
Assignment
UB
Rerun for
each scenario
42
This is a reminder of the FTA process for a trip-based 4-step model. Current news starts
modeling requires that the number of trips and their locations be fixed across alternatives and
that highway skim values remain the same across alternatives. This requires adapting the activity
based model so that outputs from the baseline model run can be used across all build alternatives
allowing us to isolate mode choice differences. New starts applications should reuse random
numbers generated in the baseline case for all build alternatives so that change is only
attributable to the differences in utilities.
1058
Page 43
Fixed
Time of day
Tour primary destination
Entire-tour mode choice
UB
Stop location
Rerun
Trip mode choice
UB
Assignment
43
User benefits for new starts applications should be tabulated at the tour and trip levels. At the
tour level individual trips are adequately represented directly when there is only one stop on the
tour but with more than one stop there may be individual trips that are not adequately represented
at the tour level in terms of user benefits. Also some models provide more detail at the trip level
in terms of transit access or sub modes and these differences need to be recognized when
summarizing user benefits at the tour level.
Once the user benefits have been calculated at the tour level, any additional user benefits that
accumulate at the trip level can be assessed. It is important to know that if the build scenario
results in a change in tour mode then the trip mode alternatives will be different and the outcome
may be counter intuitive. For example improved transit service may result in a switch from auto
to transit at the tour level and then auto is no longer included in the log-sum calculations of the
build alternative and there would be a reduction in user benefits at the trip level even though
transit service has been improved. It can be difficult to hold destination locations fixed in an
activity based model when the intermediate stops are determined after the tour mode choice
model so this needs to be accounted for in tracking user benefits at the trip level.
1059
1060
Page 44
44
Here is an example of transit user benefits assessed in San Francisco for the new central subway
project. This table shows the user benefits for the baseline scenario and the new central subway
scenario by trip purpose separately for tours and trips. Most of the user benefits are accumulated
at the trip level, but the tour benefits account for about 15% of total user benefits to account for
some benefits that cannot be accrued at the trip level.
User benefits were negative for work based tours because FTA requires that that destinations be
held fixed between the base and the build alternative. This causes complications for work-based
sub-tours because work-based sub-tour mode choices are constrained by work tour mode
choices. So, for example, if in the baseline a worker drove to work, they would select a
destination for a work-based sub-tour assuming the availability of an auto. If in the build
scenario they switch their tour mode to transit due to a transit improvement, their work-based
sub-tour destination is fixed (and assumed availability of an auto), but may be very inaccessible
now if they no longer have an auto mode available for the work-based sub-tour. This results in
negative benefits.
1061
Page 45
45
However, certain additional consideration should be taken into account. Log-sum user benefits
cannot be directly totaled across conditional choices since upper-level choices constrain lowerlevel choices and upper-level choices already include user benefits from lower-level choices. In
this regard, some drawbacks of trip-based 4-step models should be mentioned. These drawbacks
include independent non-home-based (NHB) trip mode choice. User benefits for NHB trips are
wrong and are better to be dropped.
1062
Page 46
Fixed
Time of day
UB
Stop frequency
Stop location
Rerun
A more theoretically consistent approach was adopted for the Columbus activity-based model
where user benefits are derived at the tour mode choice level. The upper models are frozen
across all scenarios while the lower portion of the model system is rerun for each scenario. It is
assumed that the tour-level mode choice shift captures the most important part of the benefits.
1063
Page 47
Fixed
Pre-mode choice
Tour primary destination
Entire-tour mode choice
UB
Stop frequency
Stop location
Rerun
Similar approach was adopted for the New York activity-based model where user benefits are
derived at the tour mode choice level. The upper models are frozen across all scenarios while the
lower portion of the model system is rerun for each scenario. It is assumed that the tour-level
mode choice shift captures the most important part of the benefits. However, note the differences
in the model structure and especially placement of the time-of-day choice.
1064
Page 48
Aggregation
Summit requirements
OD-pair structure
Mode utilities & probabilities
(?) Individual record version
Microsimulation output
Individual tour records
Probability aggregation is trivial
(?) Utility aggregation is not trivial
48
1065
Page 49
49
Aggregation of utilities allows one to swap the original individual-record choice model output
with a quasi-aggregate model that produces the same result for SUMMIT. Summit does not
know that this is a MCSM model. For the SUMMIT program it looks like a conventional
aggregate input.
1066
Page 50
i 1,2,..., I
Known aggregate:
n 1,2,..., N
Pn i
Pi
P i
n
Unknown aggregate:
Vin
Vi
50
Lets introduce the following notation and given inputs. We have to calculate the unknown
aggregate mode utilities.
1067
Page 51
exp Vi
exp V
I
j 1
Pi
2. Logsum replication:
I
N I
ln exp V j ln exp V jn N
j 1
n 1 j 1
51
I want to replace a set of individual choices with one aggregate choice (to eliminate index n). It
is done through formulation of two logical conditions. First, we want to replicate mode
probabilities. Secondly, we want to replicate the user benefits log-sum measure.
1068
Page 52
Equivalent Transformation
1. Probability replication:
exp Vi Pi C
where
C exp V j
j
2. Logsum replication:
exp V j C exp V jn
j 1
n 1 j 1
1
N
52
The first condition sets the expression for utilities with unknown constant scale. The second
condition helps identify the scale. After the equivalent transformations and substitutions we
obtain the expression for utilities.
1069
Page 53
Unique Solution
Substituting 2 to 1:
exp Vi Pi exp V jn
n 1 j 1
1
N
or equivalently:
1
Vi ln Pi ln exp V jn
N n j 1
53
The solution is unique and easy to calculate. It is also subject to interesting behavioral
interpretations. In particular, what we created is an aggregate choice model counterpart to the
disaggregate choice model.
1070
Page 54
Technical Implementation
Re-start version of ABM model:
Freeze all tour records with fixed destinations and time-ofday for baseline
Re-run mode choice and subsequent chain of models only
54
There have been new starts user benefits analysis for San Francisco and Columbus that involved
post-processing of model results for use in FTAs SUMMIT program. Since the San Francisco
activity based modeling platform is in the DaySim family of models, this post-processing could
be used with other DaySim models. And since the Columbus model is in the CT-RAMP family
of models, this post-processing could be used with other CT-RAMP models. This is a good
example of how these model applications can be reused or adapted from place to place to the
level of effort for future new starts model applications.
1071
Page 55
One interesting example of an application of activity-based model is the Manhattan Area Pricing
Study. The pricing study included multiple options to model:
1072
Page 56
An example:
Realistic modeling of Rationing by License Plate Option
56
The activity-based model offers multiple benefits, but also requires some adaptations depending
on the pricing form. Tour-level evaluation of daily fees on mode and destination choice require
the added demand / service requirements for transit; measures of educed congestion on
highways; and measures of reduced travel due to changes in destination (and stop) choices.
Disaggregate reporting of record-based results allows one to analyze who benefits and who pays
because the model keeps track of where travelers live. This is also logical and tractable for
analysis of at-work sub-tours/stops within the Central Pricing Zone (CPZ). An example that we
would like to discuss is the Realistic modeling of Rationing by License Plate Option.
1073
Page 57
Tues
Wed
Thur
Fri
57
License plate rationing (LPR) is a policy where certain cars are prohibited from driving into the
CPZ at certain hours and days based on the license plate number. An example is shown on this
slide.
1074
Page 58
The impact of License Plate Rationing is one in which 20% (or 10%) of drivers are prohibited
from driving in the CBD on certain days, based on last digit of license. LPR has affects travel
choices such that there is no impact on the primary destination choice (usually work), but there
are definite impacts on mode and intermediate stop location choices.
In modeling, we must account for opportunities to reduce the impact of ban, such as changing the
day of trip and vehicle availability within household. The household auto availability model is
the key component, because it determines whether vehicles are available for destinations in the
CPZ. In addition, car sufficiency need to be revised to reflect the number of autos minus the
number of workers.
1075
Page 59
Applied Approach
Initial calibrated trip tables
Fixed
Tour Destination
Households,
Journey generation
Auto ownership
TOD distribution
Auto Sufficiency
Auto Sufficiency for CBD
Fixed
Final assignments
59
The applied approach takes a full advantage of the micro-simulation structure and would not be
possible with an aggregate 4-step model. Essentially, the car ownership sub-model adjusts car
sufficiency for the corresponding tours and trips based on the LPR percentage. This affects all
subsequent choices for the affected tours (mode, time of day, stop frequency, etc.).
1076
Page 60
HH# Wkrs
1
2
3
4
5
6
2
1
1
1
2
2
Autos Car
Suff
3
1
2
1
4
2
a1
a2
a3
a4
1
0
1
0
2
0
60
This example illustrates how the adjustments are made for individual records. Note that car
sufficiency is calculated as a difference between the number of cars and number of workers.
Some of the cars are randomly tagged as banned based on the LPR percentage.
1077
Page 61
CBD
Other
SOV
HOV
No Change
61
Consider a household with two workers and two cars when one of the cars is subject to LPR.
This would affect probabilities of the corresponding person to travel by SOV or HOV quite
negatively (SOV might be reduced to practically zero).
1078
Page 62
CBD
Other
SOV
HOV
No Change
62
In this example, the first person would probably have to switch mode to HOV or transit,
depending on their availability.
1079
Page 63
63
Lets put this discussion into a broader contest of pricing studies. A very important aspect that
cannot be really addressed with a 4-step model is Accounting for Tolls in both Directions by
time-of-day. Scenarios to model include time-of-day-specific tolls differentiated by directions.
Required model sensitivities can be formulated as, Travelers have to see both tolls that affect:
1080
Page 64
Realistic Example
6-10 AM: $6
3-7 PM: $2
Outside
of CBD
CBD
6-10 AM: $3
3-7 PM: $5
64
Consider a realistic example where tolls are set by directions as shown in the slide. Tolls are high
for travel to and from the Congestion Pricing Zone (CBD) in the corresponding peak periods.
1081
Page 65
Inbound time
Toll, $
6-10AM ($3)
10AM-3PM
3-7PM ($5)
6-10AM ($6)
6-10AM ($3)
6-10AM ($6)
10AM-3PM
6-10AM ($6)
3-7PM ($5)
11
6-10AM ($6)
10AM-3PM
10AM-3PM
10AM-3PM
3-7PM ($5)
10AM-3PM
3-7PM ($2)
3-7PM ($5)
3-7PM ($2)
3
5
65
This setting of tolls results in a wide range of toll values paid by commuters depending on the
time-of-day choice combinations as you can see on the slide. Tolls range from zero up to $11.
How can we make this complicated reality modeled?
1082
Page 66
66
This is not simple but the activity-based model framework allows for much better behavioral
realism. With 4-step model, it is impossible to ensure any reasonable level of consistency across
trip distribution, mode choice, and time of day choice. With tour-based activity-based model, it is
still difficult to ensure a full consistency, but a much better job can be done.
1083
Page 67
Signal synchronization
Information provision
Incident response
Ramp metering
67
Conformity analysis is highly dependent on temporal traffic conditions and this additional level
of detail within activity based models is more effective at analyzing certain mitigation strategies
and policies. Many activity based models generalize these time periods when using an aggregate
trip assignment methodology and so may lose some of the advantages for temporal resolution. If
the activity based model is integrated with a traffic simulation model then these advantages can
be retained and more accurate impacts of capacity enhancement projects will be available for
conformity analysis. Some of the projects that may be difficult to quantify for conformity
without this linkage to traffic simulation include ramp metering and incident response strategies.
1084
Page 68
68
One drawback of trip based models for conformity analysis has always been the inability to trace
emissions back to the source. Activity based models allow the accumulation of the emissions at
the household level in addition to reporting at the link level. It is important to report emissions
for different traveler types or different values of time because this is how policies could affect
traveler behavior. Another useful aspect of activity based models for tracking an
Some of the more important benefits to activity-based models are the additional performance
metrics that one can produce. In some ways, regional transportation planning practices are
driving this need for additional performance metrics. We have already mentioned some of these,
and provided examples, but these are the metrics that are more commonly being required for
RTPs. They include equity measures, often across household income categories but sometimes
along other dimensions, emissions at the household level for policies around greenhouse gases
and other pollutants, energy use of vehicles and buildings, user benefits for economic sectors,
and induced travel.
1085
Page 71
Model Performance
Model design drives computational performanceuser
requirements for computational performance constrain
model design
Custom software is needed to implement demand
components of activity-based models
General purpose commercially available travel demand
modeling software is not designed to handle all of the special
decision structures, data pathways and accessibility variable
calculations, and not designed to forecast using simulation
methods
resources for hardware. There are activity based model applications running on single computers
with reasonable computational performance and other activity based model applications that
require significant hardware investments to achieve reasonable computational performance.
Computational performance is directly tied to the size of the population for the region and the
custom software being applied.
1087
Page 72
72
There are many different activity based model custom application programs developed for
specific projects. Some of these have grown as subsequent projects have built upon earlier
projects to produce more robust platforms. Many of these custom application programs are open
source, and therefore freely available, but code may be difficult to interpret without developer
assistance. Users of these platforms have contributed to the development of the software through
individual projects and benefit from software improvements funded by other agencies if they
choose to upgrade. Users are beginning to collaborate to take advantage of synergies for software
development.
1088
Page 73
SACOG (Sacramento)
NFTPO (Jacksonville)
FDOT7 (Tampa)
PSRC (Seattle)
Fresno COG
San Joaquin, Merced and Stanislaus
DVRPC (Philadelphia)
MORPC (Columbus)
TMPO (Lake Tahoe)
ARC (Atlanta)
SANDAG (San Diego)
MTC (San Francisco)
MAG/PAG (Phoenix, Tucson)
CMAP (Chicago)
SFRPC (Miami)
73
1089
Page 74
ALBATROSS (Eindhoven
University)
Applications in Netherlands
DASH (RSG)
Metro (Portland, OR)
There are also teams of academic researchers that have branded their products and implemented
or are now attempting implementations in select locations. These include custom software
written by DRCOG and CS in Denver for the FOCUS model, SimAGENT developed by U. of
Texas-Austin, U.C.-Santa Barbara, and Arizona State U. for SCAG in Los Angeles; SimTravel
developed by ASU, UC Berkeley, and U. of Arizona for a FHWA research project with a case
study in Phoenix; MatSim developed by ETH in Zrich and TU-Berlin for the Switzerland
national model and several European cities; the ILUTE model developed by the University of
Toronto and for Metrolinx; the Albatross model developed by Eindhoven University for
applications in the Netherlands; and the DASH model developed by Portland State University
and RSG for Portland Metro. These software applications demonstrate a wide variety of activity
based modeling methodologies.
1090
Page 75
User Productivity
Requires development of functionality similar to existing tripbased model software
User-friendly GUI
Scenario management
Efficient storage and organization of input and output files
Customizable outputs
Links to GIS
Links to commercially available trip-based packages for network
assignment
Data visualization
Acceptable run times
Comprehensibility and documentation
Error checking
User support
75
In addition to the custom software applications for activity based models, there are a series of
additional software tools needed to provide functionality for the travel demand forecasting which
system which is similar to utility tools used for existing trip-based model software. Even if these
utility tools exist for trip based models, they require adaptation to integrate them with activity
based modeling software. These are the types of user friendly functionality that users have come
to expect with travel demand forecasting models.
1091
Page 76
76
Other factors influencing computational performance are the temporal and spatial resolutions and
the complexity of the data and modeling structures. These structure and level of detail tradeoffs
are determined during model design and not easily adjusted once the model is complete.
Fortunately, software optimization continues to improve computational performance. Distributed
processing and multi-threading are computational methods to reduce run time and have been
quite successful in reducing run times for activity based models.
1092
Page 77
MORPC
SACOG
1,760,000
3
Intra-household Interactions
Spatial Resolution
Run Times (hours)
Yes
Yes
No
No
Zones
Zones
Parcels
Parcels
Without distribution/threading
Households per hour per iteration
Households per hour per iteration per
processor
With distribution/threading
Households per hour per iteration
Households per hour per iteration per
processor
Specifications
Hardware without distribution
Hardware with distribution
Software
610,774
3
SACOG
Households
Number of Global Iterations
146
96,154
4,521
16
36
50,898
437,500
41,250
8 processors, 16GB
RAM, 1 computer
24 processors, 48GB
RAM, 3 computers
CT-RAMP
700,000
10
33
384,615
36,164
16
330,000
1,250,000
10
109,375
4 processors, 12 GB
RAM, 1 computer
3 computers
CT-RAMP
DaySim
4 processors, 3.2 GB
RAM, 1 computer
DaySim
77
Intra-household interactions add a lot to run times (SACOG 96k households per hour per
iteration per processor compared to ARC with 4.5k). This doesnt completely separate other
software differences or the fact that SACOG has a finer resolution than ARC (parcels vs. zones)
which requires much more programming. These statistics were derived from the CMAP strategic
plan for advanced model development (June 2010).
1093
Page 78
78
Here is another example of a model design trade-off where detailed spatial resolution is
considered. Using parcel data as the most detailed spatial resolution has advantages from a landuse planning perspective but will increase run times slightly and increase data development and
cleaning significantly. Use of micro-zones can significantly reduce data development and
cleaning time and is useful from a transportation planning perspective but complicated from a
land-use planning perspective.
1094
Page 79
Research Areas
Quantifying and controlling effects of stochastic
variation and uncertainty on model results
Data visualization
High-performance computing
79
There are several areas of research which are underway to support advances in activity based
models. One area that we have already discussed is on quantifying and controlling effects of
stochastic variation and uncertainty in model results. There are two other research areas, one on
data visualization, and another on high-performance computing, that I will discuss.
1095
Page 80
80
There are several challenges to assessing stochastic variation on model results: developing
confidence intervals for outcome variables, quantifying stochastic effects of input variables, and
communicating stochastic effects to decision-makers and other stakeholders. The primary
challenge here is not one of statistics but rather one of interpretation and understanding. Many
stakeholders may not understand how to interpret model results presented in a range, especially
when comparing different alternatives.
1096
Page 81
81
Here is an example of the measure of stochastic variability and confidence intervals for the
Albatross model in Rotterdam (Netherlands). In this example, stochastic variability is quite small
and 25 to 30 runs are sufficient. Confidence intervals, however, are much higher and require
more than 100 runs to produce sufficient results.
1097
Page 82
Data Visualization
Dashboards for scenario analysis
Example from Atlanta Regional Commission (ARC)
82
Given the abundance of model outputs, data visualization seems to be limited only by
imagination and technology. For example, Atlanta has created a dashboard for scenario analysis
to review model results. Atlanta has also been testing cloud computing and the benefits to the
model user in terms of runtime and convenience.
1098
Page 83
83
Here is an example of the Atlanta data visualization dashboard. This dashboard contains four
parts:
The dashboard combines flexibility with ease of interpretation to provide meaningful results.
1099
Page 84
84
This example shows a very detailed analysis of mode share by different person types. This type
of summary based on individual (person-level) attributes is not possible with an aggregate 4-step
model.
1100
Page 85
Travelers by Age
85
This is another example of a unique angle of analysis, the distribution of travelers by age for by
mode. It is especially important and interesting to track for transit users and compare to the onboard transit surveys.
1101
Page 86
86
This is dynamic view on location of the regional population in Atlanta by hours of the day. Note
that this is not a DTA representation of traffic flows, but rather persons located at activity sites.
1102
Page 87
87
This diagram shows those who experience longest delays relative to free-flow time. Not
surprisingly, this would be workers who do the most traveling and are traveling during peak
demand periods.
1103
Page 88
Household 21
Person 1
Tour 1
HHOLD PERSON ACTIVITY PURPOSE PRIORITY START END DURATION
21
1
2114110
0
9
0
2:49
2:48
21
1
211411
3
9
3:08 7:37
4:28
21
1
211412
6
9
7:45 10:20
2:35
21
1
211421
3
9
10:25 10:29
0:04
21
1
211422
0
9
10:48 12:23
1:35
MODE
1
2
2
2
2
Start at home
TRIP 1
TRIP 2
TRIP 3
TRIP 4
Travel Time
0:19
0:08
0:05
0:19
88
This example is from the Jacksonville SHRP2 C10 project. It shows traces of a tour in DaySim
and Transims. This tour starts at home then travels to three different destinations. In this
example, we track individual vehicles, but not people riding transit or walking. Nonetheless, the
level of detail of the tour records provides insights on individual travel characteristics as well as
supply-side constraints during this period.
1104
Page 89
High-performance Computing
Advances in efficient data structures and algorithms (software
engineering)
Multi-threading
Options for distributed processing
In-house LAN serverslarge hardware investment, but local control
Leased time on remotely hosted networks (Argonne Labs)more
processors, no sunk cost in extra hardware; lack control over processing
availability, software maintenance
Cloud computingfarming out processing tasks while software resides
locally; may be public (enterprise-wide) or private (Google, Amazon)
more processors, no sunk cost in extra hardware
89
High-performance computing is rapidly changing in our world and in our industry. Advances in
efficient data structures and algorithms have opened doors for activity-based models and
continue to evolve and improve. Multi-threading and distributed processing are becoming
increasing important to achieve reasonable run times for activity-based models and provide more
opportunities for integrating Activity-based models with land use and traffic/transit microsimulation models. The options for distributed processing allow for local control on in-house
servers, leased time on remote servers, or cloud computing where processing tasks are farmed
out to public or private networks.
1105
Page 90
Summary
Disaggregate representation of individuals provides
summarization by any available attributes, enabling more
sophisticated, higher-resolution analyses of transportation
policy and investment alternatives
Application of ABMs to alternatives analysis presents some
challenges in how to present and interpret results vis--vis
the expectations of policy makers and other agencies used
to seeing trip-based model results
90
Activity-based models offer the potential for a much richer array of analysis types, depth and
sophistication than previous trip-based methods. This is largely due to the disaggregate nature of
the population and accounting for individual behavior, thus allowing for a better representation
of smaller market segments, greater flexibility in reporting, and a wider array of performance
metrics to inform decision-making. The applications of activity-based models are useful to
improve our understanding of the transportation system. The stochastic nature of AB models
requires some additional consideration on how to present and interpret a range of outcomes
rather than a single outcome.
1106
Page 91
Summary
Enhanced data visualization methods are being
developed to take advantage of this richer information
Some standardization has begun to emerge across
projects that follow the same design approach
Research into distributed computing environments
offers promise of greater performance for lower cost
91
The mechanics and interfaces for activity-based models are improving with use and are expected
to continue to get faster, better and cheaper as they become more common. The data
visualization completed to date for these models have just touched the surface of the possibilities
to mine and display the richer data contained in activity-based models. The coalescence of model
designs around two approaches in the US has created software platforms that continue to evolve
and improve and user communities that can work together for common goals. This type of
sharing of information (which has been supported by FHWA in these webinars) will benefit
users as well as those considering developing an activity-based model.
1107
Page 92
92
1108
1109
John: You are typically choosing a primary purpose and destination of the tour, if you choose a
zone or parcel in a shopping center, accessibility to other shopping opportunities is going to be
very high, so youre likely to generate one or additional trips near that location. Thats how it
would evolve in the output, creating a cluster of trips in the area.
Has anyone used the results of an activity-based model for travel time reliability analysis?
Peter: Yes. It so far has not been part of the routine for what we do, but we have developed these
methods and reported them in a research project. Travel time reliability is a very important
factor. One method weve put in practice is to generate a series of scenario type tables rather than
one, and see distributions of travel time and calculate standard deviations. The New York model
currently has this in place.
1110