Sunteți pe pagina 1din 80

Week Output

10 Analysis
TI141317 – Simulasi Sistem Industri
Simulation Procedure

Define objective, scope, and requirements

Collect and analyze system data

Build model

Validate model

Conduct experiment

Present result

Lab. KOI Jurusan Teknik Industri - ITS


Outline

•  Important of Model Verification and Validation


•  Model Verification
•  Model Validation
•  Model Building, Verification and Validation
•  Validation Examples

Lab. KOI Jurusan Teknik Industri - ITS


Importance of Model Verification and Validation

§  Model building is, by nature, very error prone


•  There is plenty of room for making errors in the
process of translation from real-world system into a
conceptual model which then translated again into
a simulation model.
§  Verification and validation processes can reduce or
eliminate these errors.

Lab. KOI Jurusan Teknik Industri - ITS


Important of Model Verification and Validation

§  Model Verification is the process of determining whether


the simulation model correctly reflects the conceptual
model.
§  Model Validation is the process of determining whether the
conceptual model correctly reflects the real system
§  Model Verification and Validation are critical to the success
of simulation project.

Lab. KOI Jurusan Teknik Industri - ITS


Illustration

Lab. KOI Jurusan Teknik Industri - ITS


Reason for Neglect

•  Despite their importance, seldom is adequate


attention given to verification and validation when
doing simulation. The primary reasons for
neglecting this activity are:
–  Time and budget pressures
–  Laziness
–  Overconfidence
–  Ignorance

Lab. KOI Jurusan Teknik Industri - ITS


Objectives of Verification and Validation

§  To produce a representative model of the system under


study
§  To increase the model credibility
§  To gradually refine the model during the development
process

Lab. KOI Jurusan Teknik Industri - ITS


Model Verification

§  Model Verification is the process of determining


whether the model operates as intended (it runs
correctly).
§  Involves the modeler more than the customer.
§  Building the model right
§  Tries to detect unintended errors in the model data and
logic and removes them.
§  Verification is the process of debugging the model è A
verified model is bug-free model

Lab. KOI Jurusan Teknik Industri - ITS


Model Verification

l  Errors or bugs in a simulation model are of


two types:
–  Syntax error : like grammatical errors and include the
unintentional addition, ommition, or misplacement of notation
that either prevent the model from running or causes it to
run incorrectly
–  Semantic errors : associated with the meaning or intention of
the modeler and are therefore difficult to detect. Often
they are logical errors.

Lab. KOI Jurusan Teknik Industri - ITS


Model Verification

Preventive Action
–  Get it right the first time (models with no bugs the
first time)
–  In practice, this isn’t always possible as bugs are
often sneaky and difficult to prevent

Practices minimizing bugs are:


–  BE CAREFUL

Lab. KOI Jurusan Teknik Industri - ITS


Model Verification

Preventive Action
To avoid errors or bugs we can use five basic principles of
structured programming
–  Top Down design, the simulation begins at a high level
–  Modularity. The model is built in modules or logical divisions
to simplify model development.
–  Compact modules. Modules should be kept as short and
simple as possible.
–  Stepwise refinement. The model is built with complexity
being progressively added.
–  Structured control

Lab. KOI Jurusan Teknik Industri - ITS


Verification Techniques
•  Conduct model code reviews
•  Check the output for reasonableness
Can be done by replacing random times and probabilistic
outcomes with constant times and deterministic outcomes.
•  Watch the animation for correct behavior
Common symptoms can be found by watching animation:
–  A terminating simulation that should have emptied the
system at the end of the simulation leaves some entities in
queue
–  A resource sits idle when there is plenty of works waiting to
be done
•  Use the trace and debug facilities provided with the software
Lab. KOI Jurusan Teknik Industri - ITS
Model Validation

l  Model Validation is the process of determining


whether the model is meaningful and accurate
representation of the real system (Hoover and
Perry, 1990)
l  Building the right model
l  Stakeholders and customers should become
heavily involved in the validation process
l  Can be very time consuming

Lab. KOI Jurusan Teknik Industri - ITS


Model Validation

•  Validation actually begins with the data-gathering


stage of a simulation project before a model is
even built. It is the garbage-in, garbage-out
syndrom that should be avoided.
•  Data validation is a matter of verifying that the
sources of information are reliable and that source
data have been correctly interpreted.
•  Validation is to be seen against the intended use of
the model and not in absolute sense.

Lab. KOI Jurusan Teknik Industri - ITS


Determining Model Validity

It is common to use combination of techniques when


validating the model:
–  Watching the animation
–  Comparing with the actual system
–  Comparing with other models
–  Conducting degeneracy and extreme condition tests
–  Checking for face validity
–  Testing againts historical data
–  Performing sensitivity analysis
–  Running traces

Lab. KOI Jurusan Teknik Industri - ITS


Steps in Validation

1.  Build a model with high face validity.


2.  Validate model assumptions.
3.  Compare the model’s input-output
transformations against those in the real system.

Lab. KOI Jurusan Teknik Industri - ITS


Determining model validity

-  A common method of validating a model of an existing


system is to compare the model performance with that
of the actual system.
-  If sufficient data are available on a performance
measure of the real system, a statictical test called
Student’s t test can be used. An F-test can be
performed to test the equality of variances of the real
system and the simulation model.

Lab. KOI Jurusan Teknik Industri - ITS


Black Box Validation

-  The inner workings of both real system and model are


unknown
-  The intention is to analyse any function of both so as to
decide whether their functioning is sufficiently similar.
-  If the model is valid in a black box sense, then the
observations of the model should be indistinguishable from
those of the real system.
-  Most commonly, the comparison is performed using statistical
methods (statistical inference, hypothesis testing)
-  Validation errors may occur as those found in hypothesis
testing so called Type I and Type II errors.

Lab. KOI Jurusan Teknik Industri - ITS


White Box Validation
§  The assumption is that the internal structures of both model and
“real” system are well understood.
§  If the stress in the black box validation is on the predictive power
of the model as captured in hypothesis tests, the stress in white
box validation is on the detailed internal structure of the model.
§  Detailed internal structure will need to focus on at least:
§  Input Distributions
§  Discrete or continuous variables
§  Infinite or finite range
§  Etc.
§  Static Logic
§  If then condition.
§  Dynamic Logic
§  Important variables and the system state can be monitored as the
program runs and this can make it relatively easy to spot errors in the
dynamic logic of the model
Lab. KOI Jurusan Teknik Industri - ITS
Optimum Level of Validation
The purpose of validation is to mitigate the risk associated
with making decision based on the model.

Optimum effort

Total cost
Cost

Validation Cost

Risk cost

Validation effort

Lab. KOI Jurusan Teknik Industri - ITS


Model Building, Verification and Validation

Steps in Model Building


1.  Observe the real system
2.  Construct conceptual model and perform conceptual
validation
3.  Translate conceptual model into a computer model
and perform verification
4.  Calibrate, verify and validate at every step

Lab. KOI Jurusan Teknik Industri - ITS


Verification and Validation

l  Verification
–  Building the model right
l  Validation
–  Building the right model
l  Verification and Validation must be conducted
simultaneously throughout the model
development process

Lab. KOI Jurusan Teknik Industri - ITS


Simulation Procedure

Define objective, scope, and requirements

Collect and analyze system data

Build model

Validate model

Conduct experiment

Present result

Lab. KOI Jurusan Teknik Industri - ITS


Approach to Simulation Output Analysis

l  Judgmental Approach

l  Statistical Approach
It depends strongly on the nature of problem,
importance of the decision, and the validity of
the input data

Lab. KOI Jurusan Teknik Industri - ITS


Judgmental Approach

l  Adequate if we are trying to find out whether


a system is capable of meeting a minimum
performance level.
l  For example: finding out whether a single
machine or a single server agent is adequate
to handle a given workload.
l  If the model relies heavily on assumptions, it
is of little value to be extremely precision in
estimating model performance.

Lab. KOI Jurusan Teknik Industri - ITS


Introduction to Statistical Analysis for
Simulation Output

•  We need statistical analysis for simulation output


because the simulation result is based on random event.
•  Model’s outputs are only estimates and NOT PRECISE
MEASURES
•  Each replication output can be considered as a sample
from all possible replication output population (This is an
infinite population, why?)
•  If one replication is independent from other replications,
then we can employ the inferential statistics. (Refer again
the issue of random number generator!)

Lab. KOI Jurusan Teknik Industri - ITS


Terminating vs Nonterminating Simulation

•  A terminating simulation is conducted when we are


interested in the behavior of the system over a
particular period or of the system that has specific
operation time period (e.g., bank operation is from
8.00 am to 3 pm).
•  A nonterminating simulation is conducted when we
are interested in the steady-state behavior of a
system or of the system that has no natural
stopping condition.

Lab. KOI Jurusan Teknik Industri - ITS


Terminating Simulations

There are three important questions to answer


in running the experiment:
1.  What should be the initial state of the
model?
2.  What is the terminating event or time?
3.  How many replications should be made?

Lab. KOI Jurusan Teknik Industri - ITS


Terminating Simulations

•  Starts at a defined state or time and ends when it reaches


some other defined state or time.
•  Common type of terminating system is one that starts
empty, runs for a while, and then empties out again before
shutting down. Example: job shops, department stores,
restaurants, banks, hair salons, and post offices.
•  Not intended to measure the steady-state behaviour of a
system although there may be periods during terminating
simulation when the system is in a steady state.

Lab. KOI Jurusan Teknik Industri - ITS


Lab. KOI Jurusan Teknik Industri - ITS
Selecting the Initial Model State

l  The initial state of the model represents how


the systems looks (or is initialized) at the
beginning of the simulation.
l  All terminating simulations do not
necessarily begin with the system being
empty and idle. For example:
- to measure the time that it takes for everyone to exit an
arena after a sporting event

Lab. KOI Jurusan Teknik Industri - ITS


Selecting a Terminating Event

l  Terminating event is an event that occurs during


the simulation that causes the simulation run to
end.
l  Terminating event might occur when a certain
time of day is reached or when certain conditions
are met.
l  If the terminating event is based on the
satisfaction of conditions other than time, then
the time that the simulation will end is not known
in advance.

Lab. KOI Jurusan Teknik Industri - ITS


Determining the Number of Replication

§  The answer to the question of how many


replications are necessary is usually based on the
analyst’s desired half-width of a confidence
interval.
§  As a general guideline, begin by making 10
independent replications of the simulation and add
more replications until the desired confidence
interval half-width is reached

Lab. KOI Jurusan Teknik Industri - ITS


Analysis of Terminating Simulation Output

Output of a bank simulation with 5 tellers, interarrival times


E(1) and service times E(4)
Number Av. delay in Av. Queue
Replication
served queue Length
1 484 1.53 1.52
2 475 1.66 1.62
3 484 1.24 1.23
4 483 2.34 2.34
5 455 2.00 1.89
6 461 1.69 1.56
7 451 2.69 2.50
8 486 2.86 2.83
9 502 1.70 1.74
10 475 2.60 2.50

Lab. KOI Jurusan Teknik Industri - ITS


Estimating Parameter

l  We can apply statistical methods to estimate the


expected or mean value of the model’s output response
l  There are two types of estimates:
1.  Point estimates (mean and standard deviation)
2.  Interval estimates (confidence intervals)
l  This is not to say that mean and standard deviation are
the only parameter estimates in which we may be
interested

Lab. KOI Jurusan Teknik Industri - ITS


Point Estimates

l  A single value estimate of a parameter of interest.


l  Point estimates are calculated for the mean and
standard deviation of the population.
l  To estimate the mean of population (denote as
µ), we simply calculate the average of the sample
values (denote as x ).

∑x i
i
x=
n

Lab. KOI Jurusan Teknik Industri - ITS


Point Estimates
l  The standard deviation for the population (denote as σ),
which is a measure of the spread of data values in the
population, is similarly estimated by calculation a
standard deviation of the sample of values (denote as s).

n 2

s=
∑ [x − x ]
i=1 i

n −1

l  The sample variance s2, used to estimate the variance of


the population σ2, is obtained by squaring the sample
standard deviation.

Lab. KOI Jurusan Teknik Industri - ITS


Point Estimates

l  For example,
Buddy opens his barbershop at 8:00 A.M. and closes at
noon on Saturday. In order to determine the exact value
for the true average number of customers getting a
haircut on Saturday morning (µ), we decide to get an
estimate of the true mean µ by spending the next 12
Saturday mornings watching TV at Buddy’s and recording
the number of customers that get a haircut between
8:00 A.M. and 12:00 noon.

Lab. KOI Jurusan Teknik Industri - ITS


Point Estimates

12
∑ x
i =1 i
x=
12
21 + 16 + 8 + ! + 10
=
12
= 13.67

12 2
[x
∑i =1 i − 13 . 67 ]
s=
12 − 1
= 4.21

Lab. KOI Jurusan Teknik Industri - ITS


Interval Estimates
l  Interval estimates constructed using x and s, on
the other hand, provide information about how far
off the point estimate x might be from the true
mean µ
l  The method used to determine this is referred to as
confidence interval estimation (a range within
which we can have a certain level of confidence
that the true mean falls)
l  The interval is symmetric about x, and the
distance that each endpoint is from x is called the
half-width (hw)

Lab. KOI Jurusan Teknik Industri - ITS


Interval Estimates
l  Confidence level is expressed as the probability P
that the unknown true mean µ lies within the interval
l  If the sample observations used to compute x and s
are independent and normally distributed, the
following equation can be used to calculate the half-
width of a confidence interval for a given level of
confidence
x ± hw
tn-1,α/2 is a factor that can be obtained from the Student’s t
table

hw =
(t n −1,α / 2 )s
n
Lab. KOI Jurusan Teknik Industri - ITS
Interval Estimates

l  α is referred to as the significance level (1–P),


may be thought of as the “risk level” or
probability that µ will fall outside the
confidence interval
l  Confidence intervals are often read as the
probability that the true but unknown mean µ
falls between the interval (x − hw) to (x + hw)
x ± hw

Lab. KOI Jurusan Teknik Industri - ITS


Interval Estimates
l  Assuming the data from the barbershop example
are independent and normally distributed, a 95
percent confidence interval is constructed as
follows:
Given: P = confidence level = 0.95
α = significance level = 1 – 0.95 =0.05
n = sample size = 12
x = 13.67 haircuts
s = 4.21 haircuts

From the Student’s t table, we find:


tn-1,α/2= t11, 0.025 = 2.021

Lab. KOI Jurusan Teknik Industri - ITS


Interval Estimates

The half-width is computed as follows:

hw =
( t11,0.025 ) s ( 2.201) 4.21
= = 2.67 haircuts
n 12
The lower and upper limits of the 95 percent
confidence interval are calculated as follows:

Lower limit = x – hw = 13.67 – 2.67 = 11.00 haircuts


Upper limit = x + hw = 13.67 + 2.67 = 16.34 haircuts

Lab. KOI Jurusan Teknik Industri - ITS


Estimating Parameter

2 2
s s (1)
x − t n −1,1−α / 2 ≤ µ ≤ x + t n −1,1−α / 2
n n
where n n
2
∑x i
i ∑ (x
i
i − x)
x= s=
n n −1
xi = the value of ith observation
n = replication number

Lab. KOI Jurusan Teknik Industri - ITS


How large n should be?

There are two approaches. The approaches are


based on:
l  absolute error (β)
l  relative error (γ)

Lab. KOI Jurusan Teknik Industri - ITS


Absolute error

l  If we are estimating µ using x so then it


is said that we have made absolute error
of β.
l  Our objective is to run the simulation
model in order to achieve the half length
less than β.
l  The number of replication that satisfy
the objective is ab ( β )
n ∗

Lab. KOI Jurusan Teknik Industri - ITS


Absolute error

l  n ( β ) is obtained from the least n



ab

that still satisfy this inequality:

2
s
t n −1,1−α / 2 ≤β (2)
n

Lab. KOI Jurusan Teknik Industri - ITS


Absolute error

Steps to calculate ab ( β ) :
n ∗

1. run the simulation for 5 or 10 replication


2. calculate s and then use it in the formula (2)
3. if the inequality has not satisfied, increase n
by 1 and then recalculate the formula (it is
assumed that s has not changed)
4. run the simulation for smallest n that satisfy
the formula (2), then we can employ
formula (1) to estimate the µ

Lab. KOI Jurusan Teknik Industri - ITS


Absolute error

l e is equal to the hw of a confidence


interval (x ± e)
l Setting e = hw and solving for n would
give an equation that could be used to
estimate the number of replications
necessary to meet a specified absolute
error amount e and significance level α
hw = e ⇒
(t n −1,α / 2 )s = e
n
2
⎡ (t n −1,α / 2 )s ⎤
n=⎢ ⎥
⎣ e ⎦
Lab. KOI Jurusan Teknik Industri - ITS
Absolute error

l  We replace tn-1,α/2 with Zα/2, which


depends only on α. The Zα/2 is from the
standard normal distribution, and its
value can be found in the last row of the
Student’s table. Note that Zα/2 = t∞,α/2,
where ∞ denotes infinity. The revised
equation is
2
⎡ (Zα / 2 )s ⎤
nʹ = ⎢ ⎥
⎣ e ⎦

Lab. KOI Jurusan Teknik Industri - ITS


Or . . .

Lab. KOI Jurusan Teknik Industri - ITS


Absolute error

l  Example:
We want to estimate the number of
replication n’ needed to be able to assert with
95 percent (1–α =0.95) confidence that the
sample mean x we compute is off by at most
2.00 haircuts (e =2.00) from the true but
unknown mean x . Based on the initial sample
of 12 observation, we know that s = 4.21
haircuts. We compute n’ as follows

Lab. KOI Jurusan Teknik Industri - ITS


Absolute error

We find Zα/2 = Z0,025 = t∞,0.025 = 1.96


Using the previous equation for n’ we obtain
2 2
⎡ ( Z 0.025 ) s ⎤ ⎡ (1.96 ) 4.21⎤
nʹ = ⎢ ⎥ =⎢ ⎥ = 17.02
⎣ e ⎦ ⎣ 2.00 ⎦
≈ 18 observations

Lab. KOI Jurusan Teknik Industri - ITS


Absolute error

We find tn-1,α/2 = t17,0.025 =


2.11. The half width is
computed as follows :

(2.11)4.28
hw =
18
= 2.13 haircuts
Increase the number of
observation (gradually) to obtain
at most 2 (of s).

Lab. KOI Jurusan Teknik Industri - ITS


Relative error

l  Relative error defines the error as a


percentage of the true parameter value.
l  Here, the objective is to satisfy the
inequality:

x−µ
≤γ
µ

Lab. KOI Jurusan Teknik Industri - ITS


Relative error

l  The number of replication needed, nrel∗


(γ ) is the
minimum n that satisfy this inequality:

2
s
t n −1,1−α / 2
n ≤γ (3)
x

l  Unfortunately, unlike s, x is unstable. So, the


formula (3) is employed iteratively

Lab. KOI Jurusan Teknik Industri - ITS


Relative error

l The iteration procedure:


0). Run n0 replication and set n = n0
1). Calculate x and δ = t n −1,1−α / 2 s 2
/n
2). If δ / x ≤ γ then use x as the point estimate
for µ and for the interval estimate, use
[x − δ , x + δ ]
Else, set n = n +1, and go to step 1

Lab. KOI Jurusan Teknik Industri - ITS


Relative error

l  It is convenient to work in terms of a relative error


(re) value (hw=re|µ).
l  This allows us to talk about the percentage error
in our point estimate in place of the absolute error
l  The relative error version of the equation become
2
⎡ ⎤
nʹ = ⎢ ( zα /2 ) s ⎥
( )
⎢ re x ⎥
⎣ (1+re) ⎦

Lab. KOI Jurusan Teknik Industri - ITS


Relative error

l  Example:
We wish to determine the approximate number of
replications needed to estimate the mean number
of haircuts given per day with an error percentage
of 17.14 percent and confidence level of 95
percent
2 2
⎡ (z )s ⎤ ⎡ (1.96)4.21 ⎤
nʹ = ⎢ 0.025 ⎥ = ⎢ ⎥ = 17.02
⎢⎣ ( 1+re ) ⎥⎦ ⎢⎣ ( 1+0.1714 )
re x 0.1714 13.67
⎥⎦
≈ 18 observations

Lab. KOI Jurusan Teknik Industri - ITS


Nonterminating Simulation

•  Does not mean that the simulation never ends, nor


does it mean that the system being simulated has no
eventual termination. It only means that the simulation
could theoretically run indefinitely with no statistical
change in behavior.
•  All observation throughout the steady-state period will
have approximately the same distribution.
•  It can change operating characteristics after a period
of time.

Lab. KOI Jurusan Teknik Industri - ITS


Analysis of Nonterminating Simulation Output

l  Nonterminating simulation is conducted is we are


interested in the steady-state behavior of the
simulation model
l  Remember, that the first customer arriving in a bank
will have no waiting time in queue. And the second
customer most likely will have little waiting time.
l  This condition does not represents the true nature of
waiting time in the bank
l  So, if we interest in long run behavior of the bank,
then we must omit the warm-up period (transient
state)

Lab. KOI Jurusan Teknik Industri - ITS


Issue for Nonterminating Simulation

l  How long the warm-up period


l  How long the simulation run should be
l  Which statistic to collect

Lab. KOI Jurusan Teknik Industri - ITS


Warm-up Period

l  Warm-up period is the transient period


l  There is no exact formula to determine when the
system (model) enter the steady-state period
l  The most common method used is the moving average
l  Here, we plot the moving average over certain
“window”. When the fluctuation of the moving average
become smoother (the average relatively constant),
then we can say that the system (model) is entering the
steady-state period

Lab. KOI Jurusan Teknik Industri - ITS


Determining The Warm Up Period

Running preliminary simulation is the easiest and most


straightforward approach. Plotting each data points
(average output response) usually helps to identify.

Lab. KOI Jurusan Teknik Industri - ITS


Average Output Response

Lab. KOI Jurusan Teknik Industri - ITS


Moving Average

l  A moving average is constructed by calculating the


arithmetic average of the w most recent data point
(averaged output responses in the data set).
l  An indicator for the end of the warm up time is
when the moving average plot appears to flatten
out (thus the routine is to begin with a small value
of w and increase it until the resulting moving
average plot begins to flatten out.

Lab. KOI Jurusan Teknik Industri - ITS


Moving Average

⎧ w
⎪ ∑ s=−w
yi+s
if i = w +1,…, m − w
⎪⎪ 2w +1
yi = ⎨ i−1


∑ y
s=−(i−1) i+s
if i = 1,…, w
⎪⎩ 2i −1

Where m denotes the total number of periods


and w denotes the window of the moving
average.

Lab. KOI Jurusan Teknik Industri - ITS


Total Average Welch
Contens for Contens per Moving
Observed Mean Queue Contens (Replications) Periods,
Periods, Average
Period i Time Totali
__
Total i
yi (6)
1 2 3 4 5 yi =
5

1 50 4.41 4.06 6.37 11.72 1.71 28.27 5.65 5.65


2 100 3.15 2.95 2.94 3.52 2.86 15.42 3.08 3.95
3 150 2.58 2.50 3.07 3.14 4.32 15.61 3.12 3.73
4 200 2.92 3.07 4.48 4.79 3.47 18.73 3.75 3.84
5 250 3.13 3.28 2.34 3.32 3.19 15.26 3.05 3.94
6 300 2.51 3.07 5.45 5.15 5.34 21.52 4.30 3.82
7 350 5.09 3.77 4.44 3.58 2.63 19.51 3.90 3.86
8 400 3.15 2.89 3.63 4.43 4.15 18.25 3.65 3.83
9 450 2.79 3.40 9.78 4.13 4.48 24.58 4.92 3.84
10 500 4.07 3.62 4.50 2.38 3.42 17.99 3.60 3.93
11 550 3.42 3.74 2.46 3.08 2.49 15.19 3.04 3.89
12 600 3.90 1.47 5.75 5.34 4.19 20.65 4.13 3.99
13 650 3.63 3.77 2.14 4.24 6.36 20.14 4.03 3.99
14 700 9.12 4.25 3.83 3.19 5.91 26.30 5.26 3.96
15 750 3.88 3.54 3.08 2.94 2.51 15.95 3.19
16 800 3.16 6.91 2.70 5.64 2.57 20.98 4.20
17 850 5.34 3.17 2.47 2.73 2.68 16.39 3.28
18 900 2.84 3.54 2.33 10.14 3.20 22.05 4.41
19 950 2.65 4.64 5.18 4.97 3.93 21.37 4.27
20 1000 3.27 4.68 3.39 2.95 3.46 17.75 3.55
Welch’s Moving Average (w=3)

Lab. KOI Jurusan Teknik Industri - ITS


Moving Average

In this example, m = 20 and w = 6

Until w = 6, we will compute using the bottom part of


the y (w)
i

Lab. KOI Jurusan Teknik Industri - ITS


Moving Average
The switch occurs at the 7th moving average because
w=6. From this point forward, we average the 2w+1
closest data point

Lab. KOI Jurusan Teknik Industri - ITS


Moving Average

l  The stopping point occurs when i = m – w. In our


case with m = 20 periods and w = 6, we stopped
when i = 20 – 6 = 14
l  Once these variables begins to exhibit steady state,
a good practice to follow would be to extend the
warm-up period by 20 to 30 percent
l  Law and Kelton (2000) recommend that w not
exceed a value greater than about m/4

Lab. KOI Jurusan Teknik Industri - ITS


Obtaining Sample Observations

1. Independent observation via replication


•  The differences between terminating and non-terminating
simulation are that
ü  the initial warm-up bias must be determined and eliminated
ü  an appropriate run length must be determined
•  One advantage of running independent replications is that
it ensure that samples are independent
•  The negative side, running through the warm-up phase for
each replication extends the length of time to perform the
replications
•  There is a possibility that the length of the warm-up
period is underestimate, causing biased result from each
replication
Lab. KOI Jurusan Teknik Industri - ITS
Obtaining Sample Observations

2.  Independent observation via batch means


•  Interval batching is a method in which a single, long run
is made and statistics are collected during separate
periods
•  Beyond the warm-up period, the run is divided into non-
overlapping intervals (interval are of an equal amount
of time or an equal number of observations)

Lab. KOI Jurusan Teknik Industri - ITS


Obtaining Sample Observations

Lab. KOI Jurusan Teknik Industri - ITS


Nonterminating Simulation Using Arena

•  How does replication in Arena works for


nonterminating simulation?

•  How do we design the modules for Batch Means of


nonterminating simulation runs?

Lab. KOI Jurusan Teknik Industri - ITS


Determining Run Length

l  Depends on:
–  Number of replication required
–  Interval between the least frequently occurring event

The choice between running independent


replications or batch intervals can be made
based on how long it takes the simulation to
reach steady state.

Lab. KOI Jurusan Teknik Industri - ITS


Determining Run Length

Lab. KOI Jurusan Teknik Industri - ITS

S-ar putea să vă placă și