Chapter 11

Output Analysis for a Single

Model
Banks, Carson, Nelson & Nicol
Discrete-Event System Simulation

Purpose

Objective: Estimate system performance via simulation

If
is the system performance, the precision of the
estimator
can be measured by:
The standard error of
.
 The width of a confidence interval (CI) for
.


Purpose of statistical analysis:

To estimate the standard error or CI .
 To figure out the number of observations required to achieve
desired error/CI.


Potential issues to overcome:

Autocorrelation, e.g. inventory cost for subsequent weeks lack
statistical independence.
 Initial conditions, e.g. inventory on hand and # of backorders at
time 0 would most likely influence the performance of week 1.


Outline

Distinguish the two types of simulation: transient vs.

steady state.
Illustrate the inherent variability in a stochastic discreteevent simulation.
Cover the statistical estimation of performance measures.
Discusses the analysis of transient simulations.
Discusses the analysis of steady-state simulations.

Type of Simulations

Terminating verses non-terminating simulations

Terminating simulation:





Runs for some duration of time TE, where E is a specified event

that stops the simulation.
Starts at time 0 under well-specified initial conditions.
Ends at the stopping time TE.
Bank example: Opens at 8:30 am (time 0) with no customers
present and 8 of the 11 teller working (initial conditions), and
closes at 4:30 pm (Time TE = 480 minutes).
The simulation analyst chooses to consider it a terminating
system because the object of interest is one days operation.

Type of Simulations

Non-terminating simulation:






Runs continuously, or at least over a very long period of time.

Examples: assembly lines that shut down infrequently, telephone
systems, hospital emergency rooms.
Initial conditions defined by the analyst.
Runs for some analyst-specified period of time TE.
Study the steady-state (long-run) properties of the system,
properties that are not influenced by the initial conditions of the
model.

Whether a simulation is considered to be terminating or

non-terminating depends on both
The objectives of the simulation study and
 The nature of the system.


Stochastic Nature of Output Data

Model output consist of one or more random variables (r.

v.) because the model is an input-output transformation
and the input variables are r.v.s.
M/G/1 queueing example:


Poisson arrival rate = 0.1 per minute;

service time ~ N( = 9.5,
=1.75).

System performance: long-run mean queue length, LQ(t).

 Suppose we run a single simulation for a total of 5,000 minutes


Divide the time interval [0, 5000) into 5 equal subintervals of 1000
minutes.
Average number of customers in queue from time (j-1)1000 to
j(1000) is Yj .

Stochastic Nature of Output Data

M/G/1 queueing example (cont.):



Batched average queue length for 3 independent replications:

Batching Interval
(minutes)
[0, 1000)
[1000, 2000)
[2000, 3000)
[3000, 4000)
[4000, 5000)
[0, 5000)

Batch, j
1
2
3
4
5

1, Y 1j
3.61
3.21
2.18
6.92
2.82
3.75

Replication
2, Y 2j
2.91
9.00
16.15
24.53
25.19
15.56

3, Y 3j
7.67
19.53
20.36
8.11
12.62
13.66

Inherent variability in stochastic simulation both within a single

replication and across different replications.
 The average across 3 replications, Y1. , Y2. , Y3. ,can be regarded as
independent observations, but averages within a replication, Y11,
, Y15, are not.


Measures of performance

Consider the estimation of a performance parameter,

(or ), of a simulated system.
Discrete time data: [Y1, Y2, , Yn], with ordinary mean:

 Continuous-time data: {Y(t), 0
t
TE} with time-weighted mean:



Point estimation for discrete time data.



The point estimator:

n
1
 =
Yi
n i =1

)

Is unbiased if E
its(expected
value is , that is if:

Is biased if:

E (
) =

Desired

Point Estimator

Point estimation for continuous-time data.



The point estimator:

[Performance Measures]

1
 =
TE

TE

Y (t )dt

Is biased in general where: E ()
 .

An unbiased or low-bias estimator is desired.

Usually, system performance measures can be put into

the common framework of
or :


e.g., the proportion of days on which sales are lost through an

out-of-stock situation, let:

1, if out of stock on day i
Y (t ) = 
0, otherwise

Point Estimator

[Performance Measures]

Performance measure that does not fit: quantile or

Pr{Y
 } = p
percentile:
Estimating quantiles: the inverse of the problem of estimating a
proportion or probability.
 Consider a histogram of the observed values Y:


Find
such that 100p% of the histogram is to the left of (smaller

than)
.

10

Confidence-Interval Estimation
[Performance Measures]

To understand confidence intervals fully, it is important to

distinguish between measures of error, and measures of
risk, e.g., confidence interval versus prediction interval.

Suppose the model is the normal distribution with mean
,

variance 2 (both unknown).


Let Yi be the average cycle time for parts produced on the ith
replication of the simulation (its mathematical expectation is
).

Average cycle time will vary from day to day, but over the long-run
the average of the averages will be close to
.
R
1
2
 Sample variance across R replications: S =
(Yi.  Y.. ) 2


R  1 i =1


11

Confidence-Interval Estimation
[Performance Measures]

Confidence Interval (CI):

A measure of error.
 Where Yi. are normally distributed.


Y.. t / 2, R
1

S
R

We cannot know for certain how far Y.. is from
but CI attempts
to bound that error.
 A CI, such as 95%, tells us how much we can trust the interval to
actually bound the error between Y.. and
.


The more replications we make, the less error there is in Y..

(converging to 0 as R goes to infinity).

12

Confidence-Interval Estimation
[Performance Measures]

Prediction Interval (PI):

A measure of risk.
 A good guess for the average cycle time on a particular day is our
estimator but it is unlikely to be exactly right.
 PI is designed to be wide enough to contain the actual average
cycle time on any particular day with high probability.
 Normal-theory prediction interval:


Y.. t / 2, R
1S 1 +

1
R

The length of PI will not go to 0 as R increases because we can

never simulate away risk.
 PIs limit is:
 z
/ 2


13

Output Analysis for Terminating Simulations

A terminating simulation: runs over a simulated time

interval [0, TE].
A common goal is to estimate:
1 n 
 = E
 Yi ,
 n i =1
 1
 = E

 TE

TE

for discrete output


Y (t )dt , for continuous output Y (t ),0
t
TE

In general, independent replications are used, each run

using a different random number stream and
independently chosen initial conditions.

14

Statistical Background
[Terminating Simulations]

Important to distinguish within-replication data from

across-replication data.
For example, simulation of a manufacturing system
Two performance measures of that system: cycle time for parts
and work in process (WIP).
 Let Yij be the cycle time for the jth part produced in the ith
replication.
 Across-replication data are formed by summarizing withinreplication data Yi. .


15

Statistical Background
[Terminating Simulations]

Across Replication:


For example: the daily cycle time averages (discrete time data)
1 R

The average:
Y.. =
Yi.
R i =1
1 R

The sample variance:
2
S =
(Yi.  Y.. ) 2


R  1 i =1
S

The confidence-interval half-width:
H = t / 2, R
1
R

Within replication:


For example: the WIP (a continuous time data)

1 TEi

The average:
Yi. =
Yi (t )dt


0
T Ei

The sample variance:

1
S =
T Ei
2
i

TEi

(Y (t )  Y ) dt
0

i.

16

Statistical Background
[Terminating Simulations]

Overall sample average, Y.. , and the interval replication

sample averages, Yi. , are always unbiased estimators of
the expected daily average cycle time or daily average
WIP.

Across-replication data are independent (different

random numbers) and identically distributed (same
model), but within-replication data do not have these
properties.

17

C.I. with Specified Precision

[Terminating Simulations]

The half-length H of a 100(1
)% confidence interval for

a mean , based on the t distribution, is given by:
H = t / 2, R
1

S
R

R is the # of
replications

S2 is the sample
variance

Suppose that an error criterion
is specified with

probability 1 -
, a sufficiently large sample size should
satisfy:

P Y..   < 
1  

18

C.I. with Specified Precision

[Terminating Simulations]

Assume that an initial sample of size R0 (independent)

replications has been observed.
Obtain an initial estimate S02 of the population variance 2.

Then, choose sample size R such that R
R0:

Since t/2, R-1
z/2, an initial estimate of R:

2
z
S


R
  / 2 0 , z / 2 is the standard normal distribution.
  
2
t
S
  / 2, R 1 0 

 R is the smallest integer satisfying R
R0 and R






Collect R - R0 additional observations.

The 100(1-
)% C.I. for :
S
Y.. t / 2, R
1
R

19

C.I. with Specified Precision

[Terminating Simulations]

Call Center Example: estimate the agents utilization  over the first 2
hours of the workday.



Initial sample of size R0 = 4 is taken and an initial estimate of the

population variance is S02 = (0.072)2 = 0.00518.
The error criterion is  = 0.04 and confidence coefficient is 1-
= 0.95,
hence, the final sample size must be at least:
2

2
 z0.025 S 0  1.96 * 0.00518
= 12.14

 =
2
0
.
04





For the final sample size:

R

t 0.025, R-1

(t / 2,R
1S0 /  )




13

14

15

2.18

2.16

2.14

15.39

15.1

14.83

R = 15 is the smallest integer satisfying the error criterion, so R - R0 = 11

additional replications are needed.
After obtaining additional outputs, half-width should be checked.

20

Quantiles

[Terminating Simulations]

In this book, a proportion or probability is treated as a special

case of a mean.
When the number of independent replications Y1, , YR is
large enough that t/2,n-1 = z/2, the confidence interval for a
probability p is often written as:

p z / 2

p (1
p )
R
1

The sample proportion

A quantile is the inverse of the probability to the probability

estimation problem:
p is given
Find  such that Pr(Y
) = p

21

Quantiles

[Terminating Simulations]

The best way is to sort the outputs and use the (R*p)th
smallest value, i.e., find
such that 100p% of the data in a
histogram of Y is to the left of
.


Example: If we have R=10 replications and we want the p = 0.8

quantile, first sort, then estimate
by the (10)(0.8) = 8th smallest
value (round if necessary).
5.6 sorted data
7.1
8.8
8.9
9.5
9.7
10.1
12.2 this is our point estimate
12.5
12.9

22

Quantiles

[Terminating Simulations]

Confidence Interval of Quantiles: An approximate (1-
)100%

confidence interval for  can be obtained by finding two values

l and u.



l cuts off 100pl% of the histogram (the Rpl smallest value of the
sorted data).
u cuts off 100pu% of the histogram (the Rpu smallest value of
the sorted data).
where pl = p
z / 2

p (1
p )
R
1

pu = p + z / 2

p (1
p )
R
1

23

Quantiles

[Terminating Simulations]

Example: Suppose R = 1000 reps, to estimate the p = 0.8

quantile with a 95% confidence interval.
First, sort the data from smallest to largest.
 Then estimate of
by the (1000)(0.8) = 800th smallest value,
and the point estimate is 212.03.
A portion of the 1000
sorted values:
 And find the confidence interval:


pl = 0.8
1.96

.8(1
.8)
= 0.78
1000
1

pu = 0.8 + 1.96

.8(1
.8)
= 0.82
1000
1

The c.i. is the 780 th and 820 th smallest values

Output
180.92
188.96
190.55
208.58
212.03
216.99
250.32
256.79
256.99

The point estimate is The 95% c.i. is [188.96, 256.79]

Rank
779
780
781
799
800
801
819
820
821

24

Output Analysis for Steady-State Simulation

Consider a single run of a simulation model to estimate a

steady-state or long-run characteristics of the system.
The single run produces observations Y1, Y2, ... (generally the
samples of an autocorrelated time series).
 Performance measure:


1 n
 = lim  Yi ,
n 
n i =1

for discrete measure

1
 = lim
TE 
TE

for continuous measure

TE

Y (t )dt ,

(with probability 1)
(with probability 1)

Independent of the initial conditions.

25

Output Analysis for Steady-State Simulation

The sample size is a design choice, with several

considerations in mind:
Any bias in the point estimator that is due to artificial or arbitrary
initial conditions (bias can be severe if run length is too short).
 Desired precision of the point estimator.
 Budget constraints on computer resources.


Notation: the estimation of
from a discrete-time output

process.
One replication (or run), the output data: Y1, Y2, Y3,
 With several replications, the output data for replication r: Yr1,
Yr2, Yr3,


26

Initialization Bias

[Steady-State Simulations]

Methods to reduce the point-estimator bias caused by using

artificial and unrealistic initial conditions:
Intelligent initialization.
 Divide simulation into an initialization phase and data-collection
phase.


Intelligent initialization
Initialize the simulation in a state that is more representative of
long-run conditions.
 If the system exists, collect data on it and use these data to specify
more nearly typical initial conditions.
 If the system can be simplified enough to make it mathematically
solvable, e.g. queueing models, solve the simplified model to find
long-run expected or most likely conditions, use that to initialize the
simulation.


27

Initialization Bias

[Steady-State Simulations]

Divide each simulation into two phases:

An initialization phase, from time 0 to time T0.
 A data-collection phase, from T0 to the stopping time T0+TE.
 The choice of T0 is important:


After T0, system should be more nearly representative of steady-state

behavior.

System has reached steady state: the probability distribution of the

system state is close to the steady-state probability distribution
(bias of response variable is negligible).

28

Initialization Bias

[Steady-State Simulations]

M/G/1 queueing example: A total of 10 independent

replications were made.

Each replication beginning in the empty and idle state.

 Simulation run length on each replication was T0+TE = 15,000
minutes.
 Response variable: queue length, LQ(t,r) (at time t of the rth
replication).
 Batching intervals of 1,000 minutes, batch means


Ensemble averages:
To identify trend in the data due to initialization bias
 The average corresponding batch means across replications:
R


Y. j =

1
Yrj


R r =1

R replications

The preferred method to determine deletion point.

29

Initialization Bias

[Steady-State Simulations]

A plot of the ensemble averages, Y ..(n, d ) , versus 1000j, for j =

1,2, ,15.

Illustrates the downward bias of the initial observations.

30

Initialization Bias

[Steady-State Simulations]

Cumulative average sample mean (after deleting d

n
1
observations):
Y.. (n, d ) =
Y. j


n  d j = d +1


Not recommended to determine the initialization phase.

It is apparent that downward bias is present and this bias can be

reduced by deletion of one or more observations.
31

Initialization Bias

[Steady-State Simulations]

No widely accepted, objective and proven technique to

guide how much data to delete to reduce initialization
bias to a negligible level.
Plots can, at times, be misleading but they are still
recommended.






Ensemble averages reveal a smoother and more precise trend as

the # of replications, R, increases.
Ensemble averages can be smoothed further by plotting a
moving average.
Cumulative average becomes less variable as more data are
averaged.
The more correlation present, the longer it takes for Y. j to
approach steady state.
Different performance measures could approach steady state at
different rates.
32

Error Estimation

If {Y1, , Yn} are not statistically independent, then S2/n

is a biased estimator of the true variance.


[Steady-State Simulations]

Almost always the case when {Y1, , Yn} is a sequence of

output observations from within a single replication
(autocorrelated sequence, time-series).

Suppose the point estimator
is the sample mean

n

Y =
i =1 Yi / n

Variance of Y is almost impossible to estimate.

 For system with steady state, produce an output process that is
approximately covariance stationary (after passing the transient
phase).


The covariance between two random variables in the time series

depends only on the lag (the # of observations between them).
33

Error Estimation

[Steady-State Simulations]

For a covariance stationary time series, {Y1, , Yn}:



Lag-k autocovariance is:
= cov(Y , Y ) = cov(Y , Y )

k
1 1+ k
i
i+k

Lag-k autocorrelation is:

k =

k
2

If a time series is covariance stationary, then the

variance
of
is:  2 
Y
n 1
k 
V (Y ) =

1 + 2
1    k 

n 
n 
k =1

The expected value of the variance estimator is:

 S2 
E   = BV (Y ),
 n 

where B =

n / c
1
n
1
34

Error Estimation

[Steady-State Simulations]



Stationary time series Yi
exhibiting positive
autocorrelation.

Stationary time series Yi
exhibiting negative
autocorrelation.

Nonstationary time series
with an upward trend

35

Error Estimation

[Steady-State Simulations]

The expected value of the variance estimator is:

 S2 
E   = BV (Y ),
 n 

where B =

n / c
1
and V (Y ) is the variance of Y
n
1

If Yi are independent, then S2/n is an unbiased estimator of V (Y )

 If the autocorrelation
k are primarily positive, then S2/n is biased
low as an estimator of V (Y ) .
 If the autocorrelation
k are primarily negative, then S2/n is
biased high as an estimator
.
V (Y )of


36

Replication Method

Use to estimate point-estimator variability and to construct a

confidence interval.
Approach: make R replications, initializing and deleting from each
one the same way.
Important to do a thorough job of investigating the initial-condition
bias:


[Steady-State Simulations]

Bias is not affected by the number of replications, instead, it is affected

only by deleting more data (i.e., increasing T0) or extending the length of
each run (i.e. increasing TE).

Basic raw output data {Yrj, r = 1, ..., R; j = 1, , n} is derived by:





Individual observation from within replication r.

Batch mean from within replication r of some number of discrete-time
observations.
Batch mean of a continuous-time process over time interval j.
37

Replication Method

Each replication is regarded as a single sample

for
n
estimating
. For replication r: Yr . (n, d ) = 1
Yrj
nd

The overall point estimator:

1 R
Y.. (n, d ) =
Yr . (n, d )
R r =1

[Steady-State Simulations]

and

j = d +1

E[Y.. (n, d )] =  n ,d

If d and n are chosen sufficiently large:



n,d ~
.

is an approximately unbiased estimator of
.

To estimate standard error of Y.. , the sample variance and
 Y.. ( n, d )

standard error:
1 R
1  R 2
2
2
(
)
S =
Y
Y
Y
R
Y

=






..
.. 
r.
r.
R  1 r =1
R  1  r =1

2

and

s.e.(Y.. ) =

S
R
38

Replication Method

[Steady-State Simulations]

Length of each replication (n) beyond deletion point (d):

(n - d) > 10d
Number of replications (R) should be as many as time
permits, up to about 25 replications.
For a fixed total sample size (n), as fewer data are
deleted ( d):
C.I. shifts: greater bias.
 Standard error of Y.. ( n, d ) decreases: decrease variance.


Reducing
bias

Trade off

Increasing
variance

39

Replication Method

[Steady-State Simulations]

M/G/1 queueing example:

Suppose R = 10, each of length TE = 15,000 minutes, starting at
time 0 in the empty and idle state, initialized for T0 = 2,000
minutes before data collection begins.
 Each batch means is the average number of customers in queue
for a 1,000-minute interval.
 The 1st two batch means are deleted (d = 2).
 The point estimator and standard error are:


Y.. (15,2) = 8.43



and

s.e.(Y.. (15,2) )= 1.59

The 95% C.I. for long-run mean queue length is:

Y..  t / 2, R 1S / R

Y.. + t / 2, R 1S / R
8.43  2.26(1.59)
LQ
8.42 + 2.26(1.59)

A high degree of confidence that the long-run mean queue length

is between 4.84 and 12.02 (if d and n are large enough).
40

Sample Size

[Steady-State Simulations]

To estimate a long-run performance measure, , within

with confidence 100(1-
)%.
M/G/1 queueing example (cont.):
We know: R0 = 10, d = 2 and S02 = 25.30.
 To estimate the long-run mean queue length, LQ, within  = 2
customers with 90% confidence (
= 10%).
2
 Initial estimate:
2


 z S  1.645 (25.30)
R
 0.05 0  =
= 17.1
2
2
  

Hence, at least 18 replications

are needed, next try R = 18,19,
2
using R
(t0.05, R1S 0 /  ) . We found that:
R = 19
(t0.05,191S 0 /  ) = (1.74 * 25.3 / 2) 2 = 18.93
2

Additional replications needed is R R0 = 19-10 = 9.

41

Sample Size

[Steady-State Simulations]

An alternative to increasing R is to increase total run

length T0+TE within each replication.


Approach:

Increase run length from (T0+TE) to (R/R0)(T0+TE), and

Delete additional amount of data, from time 0 to time (R/R0)T0.

Advantage: any residual bias in the point estimator should be

further reduced.
 However, it is necessary to have saved the state of the model at
time T0+TE and to be able to restart the model.


42

Batch Means for Interval Estimation

[Steady-State Simulations]

Using a single, long replication:

Problem: data are dependent so the usual estimator is biased.
 Solution: batch means.


Batch means: divide the output data from 1 replication (after

appropriate deletion) into a few large batches and then treat the
means of these batches as if they were independent.

A continuous-time process, {Y(t), T0
t
T0+TE}:



k batches of size m = TE/k, batch means:

1 jm
Y (t + T0 )dt


j
m
(

1
)
m

A discrete-time process, {Yi, i = d+1,d+2, , n}:
jm
 k batches of size m = (n d)/k, batch means:
1
Yj =
Yi + d


m i =( j 1) m +1
Yj =

43

Batch Means for Interval Estimation

[Steady-State Simulations]

Y1 , ..., Yd , Yd +1 , ..., Yd + m , Yd + m +1 , ..., Yd + 2 m , ... , Yd + ( k
1) m +1 , ..., Yd + km

deleted

Y1

Starting either with continuous-time or discrete-time data, the

variance of the sample mean is estimated by:
2

S
1
=
k
k

Yk

Y2

j =1

(Y j  Y )2 =
k 1

j =1

Y j2  kY 2
k (k  1)

If the batch size is sufficiently large, successive batch means

will be approximately independent, and the variance estimator
will be approximately unbiased.
No widely accepted and relatively simple method for choosing
an acceptable batch size m (see text for a suggested
approach). Some simulation software does it automatically.
44

Summary

Stochastic discrete-event simulation is a statistical experiment.




Distinguish: terminating simulations and steady-state simulations.

Steady-state output data are more difficult to analyze



Purpose of statistical experiment: obtain estimates of the performance

measures of the system.
Purpose of statistical analysis: acquire some assurance that these
estimates are sufficiently precise.

Decisions: initial conditions and run length

Possible solutions to bias: deletion of data and increasing run length

Statistical precision of point estimators are estimated by standarderror or confidence interval

Method of independent replications was emphasized.

45