Documente Academic
Documente Profesional
Documente Cultură
NO 5, pp 3t9-325, 1991
Printed m Great Britain
0191-2607191 $3 00 + .00
1991 Pergamon Press p|c
INTRODUCTION
320
M. BIELLI et aL
321
TP module
The Traffic Process Model module is the backbone of the system and provides for the background
knowledge concerning the ongoing traffic situation.
Most of the existing UTC systems does not embody
any explicit model of the evolving traffic, merely
ctuetors
~gulatton
CRoadslde
sensors ~
tore
'
Truffle
/
measurement~
'
Hybnd~odule
Hybrid module
=1
Volumes
Densities
Speeds
Tendencies
RunnJnc,~Signal
l PI~
Trefflc Slgne! Plen8
Network
structure
Statistical
TP-HndBla
Fig. 1. The overall architecture.
Deta
322
M. BIELLIet al.
In this sense, the TP module is intended to embody snapshots of the traffic process as it progresses
in time, reflecting, instant by instant and as close as
possible, the actual changes on the real network.
This description is based both on static and dynamic
descriptive attributes of the process, such as: (a)
structural information about the network: for example, geometry of the network, hierarchical structure,
capacity of single links, turning possibilities in each
node, additional restrictions like bus lanes, etc.; (b)
macroscopic traffic flows parameters: that is, a~:tual
volume, density, speed on links, etc., for each network element, both in values and tendencies (decreasing/increasing/stationary), acquired by sensors
or inferred (by AI-module#1); and (c) empirical
knowledge of specbCicphenomena: for example, relevance of a particular intersection during the different
time periods of the day or related to traffic composition, etc.; its attitude to propagate problems to
neighbouring network elements; what areas should
be given the highest priorities in solving congestion;
what type of traffic to expect in a given area at a
given time of the day, and so on.
On the whole, this forms an enriched picture of
the ongoing traffic, which extends the partial picture
as formed by the set of data acquired by roadside
sensors.
A l-module#1
The TP module is fed by the AI-module#1, which
provides the data completion functionality mentioned above. AI-module#1 has access to the set of
measured data coming from sensors and to the current state of the Traffic Process. Using such information as a set of constraints on the unknown d a t a that is, using the known data and the empirical
knowledge of what is likely to happen, given the
actual situation-AI-module#1 tries to infer the
missing information in order to maintain a complete
picture of the evolving traffic.
Furthermore, it classifies the situation with respect to some criteria useful for selecting the appropriate signal plan. With the rules usually employed
in present practice, exceeding a single threshold
might be sufficient to identify a different situation.
As this might not be an adequate interpretation, AImodule#1 uses more complex rules which take into
account other factors, like the time of the day, the
day of the week, weather conditions, etc., to adapt
the interpretation of traffic flows to changing conditions.
Al.module#2
Using the results of data completion and analysis,
AI-module#2 draws the conclusions in order to determine the best decision as regards the actual traffic
control action. More specifically, it has to: (a) assess
the results of analysis and diagnosis performed by
AI-module#1; (b) judge on the adequacy of the current active signal plan. In the offline calculation process that leads to different SPs, to each SP corresponds an index that assesses the capability of the SP
Model tuner
The TP module maintains a qualitative picture
of the evolving traffic. The Model Tuner is a
knowledge-based module which performs metareasoning on the Traffic Process Model. It represents,
to some extent, a learning component that compares
the traffic situation (as described by acquired data)
with the picture of the evolving traffic as elaborated
by the Traffic Process Model module. In principle,
the discrepancies between the current traffic situation and the situation predicted by the TP are used
to update the rules which are embedded in it. For
example, one can imagine a metarule, in the Model
Tuner, which encode the following knowledge: if input data suggest that it is a typical rush-hour traffic
at link L3 and the TP states that it is not, then update
the rules, for it is rush-hour traffic at link L3.
Ai METHODS AND TECHNIQUES
R ule-based programming
It is the best method, today, for realizing systems
which capture expert knowledge about some domain. Production systems are the most common
rule-based systems. A problem solving model based
on production systems consists of: (a) a knowledge
base, that is, a set of rules, each of which is formed
by a precondition and a set of actions (e.g., P --* Q
& R or ifP then Q and R) and captures some heuristic of the expert; (b) one or more data bases that
contain information about the domain and about the
problem-solving state (the working memory); (c) a
control strategy, that is, a way for controlling and
directing the reasoning steps performed by a rule
interpreter.
The rule interpreter executes a recognize-act cycle
that repeatedly looks for production rules whose
condition part matches some data in the working
memory, activates that rules (according to some selection mechanism which is part of the control strategy), and executes their action parts.
323
compded knowledge.
Different ways of representing and using knowledge are often required for different reasoning tasks,
for example, for reasoning on hypothetical consequences of actions or for making diagnoses. Modelbased reasoning is a deeper form of encoding and
using the knowledge of the domain. It implies the
use of models that describe the underlying system
structure and its components interactions. Such descriptions can be, for example, functional or causally
oriented abstraction of the system one is reasoning
about. Model-based reasoning provides for the basis
to analyse and explain the behaviour of the system
that is modelled, to capture its temporal aspects, and
SO on.
The TP module is to be encoded in this way. The
objects which make up the Traffic Process, their attributes, and their relationships must be represented.
This includes both static elements (the topology of
the network) and dynamic elements (the state of each
element in terms of descriptive variables, like volume
and density and their trends; causal relationships between descriptive variables, e.g. "the increase of traffic volume on link#n and link#m is the cause of a
precongested state on intersection#p," etc.). Qualitative modelling theories (Bobrow, 1985) are the best
tool available today for representing physical systems and perform qualitative reasoning on their behaviour.
Constraint programming
Constraint programming is a quite recent AI programming paradigm (McDermott and Charniak,
1985). It is based on techniques for representing constraints among variables and to use such constraints
for propagating values among variables (constraint
satisfaction). Problems which are tractable by constraint satisfaction techniques are characterized by:
(a) a set of elements to be interpreted, usually arranged as a network of elements, that is, a plausible
value is to be assigned to the elements; and (b) a set
of constraints among adjacent elements restricting
the value assignment with respect to the adjacent
elements.
Given some initial assignment of values to a subset of elements, a constraint satisfaction mechanism
tries to achieve an assignment of values to all elements that is consistent with the constraints. Depending on the kind of values, one can have symbolic
or numeric constraint propagation.
This method for representing and processing the
information can be used for the reconstruction of
data on nonequipped links and intersections. The set
of elements are the links and the intersections represented in the Traffic Process. The constraints represent the structural knowledge of the network, acquired traffic data, signal plans, etc. Using
constraint propagation, starting with the actual state
(a starting assignment), a consistent present traffic
situation might be inferred.
M. BIELLI el QL
324
Domain
Developers
References
INRETS, France
Field trials
(Paris)
Univ. of Taiwan
Prototype
Univ. Valencia,
ETRA (Valencia)
Spain
Incident detection and
Univ. Minnesota,
warning
USA
Safety and traffic moni- Umv. Roma, Italy
toring
Traffic control
Highways
Development
Stage
Prototype
Prototype
Lakshminarayanan &
Stephanedes (1989)
Feasibihty
study
Vtrgmia Dept. of
Prototype
Transportation,
USA
Univ. Madrid, Spain Prototype
While the research efforts devoted to the introduction of AI concepts into conventional traffic control systems are comparatively few with respect to
similar efforts in other industrial application areas,
nevertheless, as already mentioned, in the last few
years some major research work has been undertaken along with this direction. Some of such work
is being carried out in the framework of coordinated
European research programmes. To give a rapid
overview of the current research on this topic, some
of the major projects are briefly described below.
DRIVE
The DRIVE progamme started in January 1989
and is currently funding about 50 research projects
aimed at the development and the application of new
road traffic informatics systems for improving traffic efficiency and safety. A wide range of aspects
are considered under DRIVE, from traffic control to
modelling and simulation, safety, impact evaluation,
and so on. Several projects dealing with specific issues (like, e.g., incident detection or user information and warning systems) make some reference to
the use of AI and KBS concepts. Three projects are
explicitly concerned with the analysis and the development of AI methods to be applied to traffic engineering and control problems.
One of these three projects is a feasibility study
of AI potentialities with respect to road traffic and
transportation engineering and has the final objec-
Allport D. (1988) Interpreting incident reports. IEE Colloquium on Apphcation of Expert Systems In Road
Transportation, Preprints. IEE, London.
Bell M. C. (1988) The fundamental issues of an expert system for urban traffic control, lEE Colloquium on Apphcatlon of Expert Systems in Road Transportation,
9(7), 1-6. lEE, London.
Blosseville J. M., Krafft C., Lenoir F., Motyka V., and
Beucher S. (1989) TITAN: A traffic measurement system using image processing techniques. IEE 2rid Int.
Conf. on Road Traffic Monitoring, Preprints. IEE,
London.
Bobrow D. G. (1985) Qualitative Reasoning about Physical
Systems. MIT Press, Cambridge, MA.
325