LECTURE 40
SIMULATION MODELING VI
DR. ANUP KUMAR TRIPATHI
DEPARTMENT OF MINING ENGINEERING
NATIONAL INSTITUTE OF TECHNOLOGY KARNATAKA (NITK)
SURATHKAL − 575025, INDIA
1
SIMULATION MODELING VI
METHODS FOR GATHERING STATISTICAL OBSERVATIONS
Simulation is a statistical experiment, and its output must be interpreted using proper statistical
inference tools.
To accomplish this task, a simulation experiment must satisfy three conditions:
1. Observations are drawn from stationary distributions.
2. Observations are sampled from a normal population.
3. Observations are independent.
In a strict sense, the simulation experiment does not satisfy any of these conditions.
Nevertheless, we can ensure that these conditions remain statistically acceptable by restricting the
manner in which the observations are gathered.
2
First, we consider the issue of stationary distributions.
Simulation output is a function of the length of the simulated period.
The initial period produces erratic behavior and is usually referred to as the transient or warm-up
period.
When the output stabilizes, the system operates under steady state.
Unfortunately, there is no definitive way to predict the start point of steady state in advance.
In general, a longer simulation run has better chance of reaching steady state — meaning that the
problem is addressed by using a sufficiently large sample size.
3
Next, we consider the requirement that simulation observations are drawn from a normal
population.
This requirement is realized by using the central limit theorem, which confirms that the
distribution of the average of a sample is asymptotically normal regardless of the parent
population.
The central limit theorem is thus the main tool we use for satisfying the normal distribution
assumption.
4
The third condition deals with the independence of the observations.
The nature of the simulation experiment does not guarantee independence among successive
simulation observations.
However, by using sample average to represent a simulation observation, we can alleviate the
problem of lack of independence.
This is particularly true when we increase the time base used to compute the sample average.
Having discussed the peculiarities of the simulation experiment and ways to circumvent them, we
present the two most common methods for collecting observations in simulation, which are
subinterval method and replication method.
5
1. Subinterval Method
Figure 5 illustrates the idea of the subinterval method.
Suppose that the length of the simulation run is T time units.
The subinterval method first truncates an initial transient period, and then subdivides the
remainder of the simulation run into n equal subintervals (or batches).
The average of a desired measure of performance (for example: queue length or waiting time in
queue) within each subinterval is then used to represent a single observation.
Truncation of the initial transient period means that no statistical data are collected during that
period.
6
Figure 5: Collecting simulation data using the subinterval method.
7
Example 7
Figure 6 shows the change in queue length in a single-queue model as a function of the simulation
time.
The simulation run length is T = 35 hrs, and the length of the transient period is estimated to equal
5 hrs.
The time base for an observation is 6 hrs, which produces n = 5 observation.
Determine the average queue length for the simulation run.
Figure 6: Change in queue length with simulation time in the present Example.
8
Solution
Let represent the average queue length in batch i.
Because the queue length is a time-based variable, we have:
where Ai is the area under the queue-length curve associated with batch (observation) i, and t (= 6)
is the time base per batch.
The data in Figure 6 produce the following observations:
9
The sample mean and variance can be used to compute a confidence interval, if desired.
The computation of the sample variance in the present Example is based on the following familiar
formula:
10
2. Replication Method
In the replication method, each observation is represented by an independent simulation run in
which the transient period is truncated, as illustrated in Figure 7.
The computation of the observation averages for each batch is the same as in the subinterval
method.
Figure 7: Collecting simulation data using the replication method.
11
The advantage of the replication method is that each simulation run is driven by a distinct 0-1
random number stream, which yields statistically independent observations.
The disadvantage is that each observation may be biased by the initial effect of the transient
conditions.
Such a problem may be alleviated by making the run length sufficiently large.
12