Lecture Notes of

Modeling and Simulation

7th Sem IT

BCS-408
Module I

MODELING & SIMULATION (3-1-0)

Cr.-4
(10 Lectures)

Inventory Concept: The technique of Simulation. : 1 class

Major application areas, concept of a System.
: 1 class
Environment. : 1 class
Continuous and discrete systems.
: 1 class
Systems modeling, types of models. : 1 class
Progress of a Simulation Study.
: 1 class
Monte Carlo Method. : 1 class
Comparison of Simulation and Analytical Methods. : 1 class
Numerical Computation Technique for discrete and continuous models.
Continuous System Simulation. ;1 class
Revision
Module II

: 1 class

(12 Lectures)

Probability Concepts in Simulation: 2 classes

Stochastic variables, Discrete and Continuous Probability Functions.
2 classes
Numerical evaluation of continuous probability functions, continuous uniformly distributed
random numbers.
: 2 classes
Random Number Generators Linear congruential Generator, Mid Square Method,
Multiplicative Congruential generator, rejection Method. : 2 classes
Testing of random Numbers. : 2 classes
Generation of Stochastic variants. : 1 class
Arrival Patterns Service times.
: 1 class
Revision
Module III

(10 Lectures)

Discrete System Simulation and GPSS: Discrete Events, Representation of Time, generation of
arrival patterns.
: 2 classes
Fixed time step versus next event simulation, Simulation of a
Telephone System, Delayed calls . : 2 classes
Introduction to GPSS: Creating and moving transactions, queues. : 2 classes
Facilities and storages, gathering statistics, conditional transfers, program control statements,
priorities and parameters.
: 2 classes
Standard numerical attributes, functions, gates, logic switches and tests, Variables, Select and
Count. : 2 classes
Revision

Module IV

(10 Lectures)

Simulation Languages and Practical Systems

: 1 class
Continuous and discrete systems languages, factors in the section of discrete systems simulation
language.
; 2 classes
Computer model of queuing, inventory and scheduling systems. : 2 classes
Design and Evaluation of simulation Experiments: Length of simulation runs, validation,
variance reduction techniques.
: 2 classes
Experimental layout, analysis of simulation output,
Recent trends and developments.
: 1 class
Revision
Books:
1.
2.

System Simulation Geoffrey Gordon, 2nd Edition, PHI

System Simulation with Digital computer Narsingh Deo, PHI

Module-I
Objectives:

To give an overview of the course (Modeling & simulation).

Define important terminologies.
Classify systems/models

System:
any set of interrelated components acting together to achieve a common objective.
Examples:
1. Battery

Consists of anode, cathode, acid and other omponents.

These components act together to achieve one objective like preserving electricity.
2. University
Consists of professors, students and employees.
These objects act together to achieve the objective of teaching & learning process.
A system consists of

Inputs
Elements that cause changes in the systems variables.

Outputs
Response
Systems (process)
Defines the relationship between the inputs and outputs
Some Possible Inputs
Inlet flow rate

Temperature of
entering material


Concentration of
entering material
Some Possible Outputs
Level in the tank

Temperature of
material in tank
Outlet flow rate

Concentration of material in tank

Qn: What inputs and outputs are needed when we want to model the Inventory
Control System?
Model: A model describes the mathematical relationship between inputs and outputs.
Simulation: is the process of using the mathematical model to determine the
response of the system in different situations in a Computer system.
Classification of Systems
Systems can be classified based on different criteria:

Spatial characteristics: lumped & distributed

Continuity of the time

variable: Continuous, discrete-time

Quantization of dependent
variable: Quantized & Nonquantized

Parameter variation: time varying & fixed (time-invariant)

Superposition principle: linear & nonlinear

Continuous-time System:
The signal is defined for all t in an interval [ti, tf]
Discrete-time System:
The signal is defined for a finite number of time points {t0, t1,}
A system is linear:
if it satisfies the super position principle.
A system satisfies the superposition principle if the following conditions are satisfied:
1. Multiplying the input by any constant, multiplies the output by the same constant.
2. The response to several inputs applied simultaneously is the sum of individual
response to each input applied separately.

Quantized variable System:

The variable is restricted to a finite or countable number of distinct values.

Non-Quantized variable System:

The variable can assume any value within a continuous range.

Characteristics of Lumped Systems:

Only one independent variable ( t )

No dependence on the spatial coordinates

Modeled by ordinary differential equations

Needs a finite number of state variables

Distributed System:

More than one independent variable

Depends on the spatial coordinates or some of them.

Modeled by partial differential equations

Needs an infinite number of state variables.

Time variant Systems:

Observes the change of state of the variable regularly and records the
related information at that point of change.
Fixed-Time Event Systems:

Changes of the state of the variable occurs at some constant interval.

Models and types:
Models are the replica of systems which can be represented physically or
mathematically.
All the physical and mathematical models can further be divided into categories
like: static and dynamic.
Simulation models may be either deterministic or stochastic (meaning probabilistic). In a
deterministic simulation, all of the events and relationships among the variables are governed
entirely by a combination of known, but possibly complicated, rules . The advantage of
simulation is that you can still answer the question even if the model is too complicated to solve
analytically.
In a stochastic simulation, random variables are included in the model to represent the
influence of factors that are unpredictable, unknown, or beyond the scope of the model we use in
the simulation.
.
In many applications, such as a model of the tellers in a bank, it makes sense to incorporate
random variables into the model. In the case of a bank, we might wish to assume that there is a
stream of anonymous customers coming in the door at unpredictable times, rather than explicitly
modeling the behavior of each of their actual customers to determine when they plan to go to the
bank.
It is worth noting here that it is well known in statistics that when we combine the actions of a
large population of more-or-less independently operating entities (customers, etc.) the resulting
behavior appears to have been randomly produced, and that the patterns of activity followed by
the individual entities within that population are unimportant. For example, all of the telephone
systems designed in the last 60 years are based on the (quite justified) assumption that the umber

of calls made in fixed length time intervals obeys the Poisson distribution. Thus the generation
and use of random variables is an important topic in simulation.

Static Versus Dynamic Simulation Models

Another dimension along which simulation models can be classified is that of time. If the model
is used to simulate the operation of a system over a period of time, it is dynamic. The baseball
example above uses dynamic simulation. On the other hand, if no time is involved in a model, it
is static. Many gambling situations (e.g., dice, roulette) can be simulated to determine the odds
of winning or losing. Since only the number of bets made, rather than the duration of gambling,
matters, static simulation models are appropriate for them
Monte Carlo Simulation (named after a famous casino town1 in Europe) refers to the type of
simulation in which a static, approximate, and stochastic model is used for a deterministic
system.
Let us now look at an example of Monte Carlo simulation. Consider estimating the value of by
finding the approximate area of a circle with a unit radius. The first quadrant of the circle is
enclosed by a unit square.
If pairs of uniformly distributed pseudo-random values for the x and y coordinates in [0, 1] are
generated, the probability of the corresponding points falling in the quarter circle is simply the
ratio of the area of the quarter circle to that of the unit square.
The above program is an example of Monte Carlo integration, by which definite integral of
arbitrary but finite functions can be estimated. Consider the following integral:
Such an integral can be partitioned into segments above or below the horizontal axis. Without
loss of generality, let us assume f (x)0, a x b. We can then bound the curve with a rectangle
with borders of x = a, x = b, y = 0, and y = y max, where y max is the maximum value of f (x) in
the interval [a, b]. By generating random points uniformly distributed over this rectangular area,
and deciding whether they fall above or below the curve f (x), we can estimate the integral.
Dynamic Simulation
In the remainder of chapter, dynamic stochastic simulation models will be emphasized. A
dynamic simulation program that is written in a general purpose programming language (such as
Turing) must:
i) keep track of simulated time,
ii) schedule events at simulated times in the future, and
iii) cause appropriate changes in state to occur when simulated time reaches the time at which
one or more events take place.
The structure of a simulation program may be either time-driven or event-driven, depending on
the nature of the basic loop of the program. In time-driven models (see Figure 2a), each time
through the basic loop, simulated time is advanced by some small (in relation to the rates of
change of the variables in the program) fixed amount, and then each possible event type is
checked to see which, if any, events have occurred at that point in simulated time. In eventdriven models (see Figure 2b), events of various types are scheduled at future points in

simulated time. The basic program loop determines when the next scheduled event should occur,
as the minimum of the scheduled times of each possible event. Simulated time is then advanced
to exactly that event time, and the corresponding event handling routine is executed to reflect the
change of state that occurs as a result of that event.
Constructing a Simulation Model
1. Identification of Components
Our first task is to identify those system components that should be included in the model. This
choice depends not only on the real system, but also on the aspects of its behavior that we intend
to investigate. The only complete model of a system is the system itself. Any other model
includes assumptions or simplifications. Some components of a system may be left out of the
system model if their absence is not expected to alter the aspects of behavior that we wish to
observe.
We assume that the bank automation problem specification includes the following information:
i) the times between successive arrivals of customers, expressed as a probability distribution,
ii) the distribution of the number of liters of gasoline needed by customers,
iii) the distribution of how long service at the pump takes as a function of the number of litres
needed,
iv) the probability that an arriving customer will balk as a function of the number of cars already
waiting
at the service station and the number of liters of gasoline he needs,
v) the profit per liter of gasoline sold, and
vi) the cost per day of running a pump (including an attendants salary, etc.).
2. Entities
Customers, resources, service facilities, materials, service personnel, etc., are entities. Each type
of entity has a set of relevant attributes. In our service station example, the entities are cars, with
the attributes arrival time and number of liters needed, and pumps, with the attribute time to
the completion of service to the current customer.
3. Events
Events are occurrences that alter the system state. Here the events are the arrival of a customer at
the station, the start-of-service to a customer at a pump, and the completion-of-service to a
customer at a pump. The first arrival event must be scheduled in the initialization routine; the
remaining arrivals are handled by letting each invocation of the arrival routine schedule the next
arrival event. The scheduling of a start-of-service event takes place either in the arrival routine, if
there are no other cars waiting and a pump is available, or in the completion-of service routine,
otherwise. Each time the start-of-service routine is invoked, the completion-of-service at that
pump is scheduled.
Besides the types of events identified in the real system, there are two other pseudo-events that
should be included in every simulation program. End-of-simulation halts the simulation after a
certain amount of simulated time has elapsed, and initiates the final output of the statistics and
measurements gathered in the run. End-of simulation should be scheduled during the
initialization of an event-driven simulation; for time-driven simulations, it is determined by the
upper bound of the basic program loop. A progress-report event allows a summary of statistics
and measurements to be printed after specified intervals of simulated time have elapsed. These
progress-report summaries can be used to check the validity of the program during its
development, and also to show whether or not the system is settling down to some sort of
equilibrium (i.e., stable) behavior.
4. Groupings

Similar entities are often grouped in meaningful ways. Sometimes an ordering of the entities
within a group
is relevant. In the service station example, the groupings are available pumps (that are not
currently serving any auto), busy pumps (ordered by the service completion time for the
customer in service), and autos awaiting service (ordered by time of arrival). In a Turing
program, such groupings can be represented as linked lists of records, as long as a suitable link
field is included in the records for that entity type.
5. Relationships
Pairs of non-similar entities may be related. For example, a busy pump is related to the auto that
is being served. Relationships can be indicated by including in the record for one of the entities
in a pair a link to the other entity. In some cases, it may be desirable for each of the entities in the
pair to have a link to the other entity. For example, the relationship between a busy pump and the
auto being served can be represented by a link from the pump record to the corresponding auto
record.
6. Stochastic Simulation
In a stochastic simulation, certain attributes of entities or events must be chosen at random
from some probability distribution. In the service station example, these include the customer
inter arrival times (i.e., the times between successive arrivals), the number of liters needed, the
service time (based on the number of liters needed), and whether the customer decides to balk
(based on the length of the waiting queue and the number of liters needed).
7. Strategies
Typically, a simulation experiment consists of comparing several alternative approaches to
running the system to find out which one(s) maximize some measure of system performance.
In the case of the service station, the strategies consist of keeping different numbers of pumps in
service.
8. Measurements
The activity of the system will be reflected in measurements associated with entities, events,
groups, or relationships. Such measurements are used in specifying the performance of the
system. If the measurement is associated with an entity, it should be recorded in a field contained
within the record for that entity. For measurements associated with events, groupings or
relationships, additional variables or records must be declared to record the measured quantities.
In our example system, we might wish to measure the numbers of customers served, liters of
gasoline sold, customers who balk, and liters of gasoline not sold to balking customers, the total
waiting time (from which we can calculate average waiting time), the total service time (to get
pump utilization), and the total time that the waiting queue is empty.
conditions due to customers left in the system when end-of-simulation is reached are handled
properly.
Advantages and disadvantages of simulation
Competition in the computer industry has led to technological breakthroughs that are allowing
hardware companies to continually produce better products. It seems that every week another
company announces its latest release, each with more options, memory, graphics capability, and
power.
What is unique about new developments in the computer industry is that they often act as a
springboard for other related industries to follow. One industry in particular is the simulation
software industry. As computer hardware becomes more powerful, more accurate, faster, and
easier to use, simulation software does too.
The number of businesses using simulation is rapidly increasing. Many managers are realizing
the benefits of utilizing simulation for more than just the one-time remodeling of a facility.
Rather, due to advances in software, managers are incorporating simulation in their daily

operations on an increasingly regular basis.

Advantages:
For most companies, the benefits of using simulation go beyond just providing a look into the
future. These benefits are mentioned by many authors [Banks, Carson, Nelson, and Nicol, 2000;
Law and Kelton, 2000; Pegden, Shannon and Sadowski, 1995; and Schriber, 1991]
and are included in the following:
Choose Correctly. Simulation lets you test every aspect of a proposed change or addition
without committing resources to their acquisition. This is critical, because once the hard
decisions have been made, the bricks have been laid, or the material-handling systems have been
installed, changes and corrections can be extremely expensive. Simulation allows you to test
your designs without committing resources to acquisition.
Time Compression and Expansion. By compressing or expanding time simulation allows you
to speed up or slow down phenomena so that you can thoroughly investigate them. You can
examine an entire shift in a matter of minutes if you desire, or you can spend two hours
examining all the events that occurred during one minute of simulated activity.
Understand "Why?" Managers often want to know why certain phenomena occur in a real
system. With simulation, you determine the answer to the "why" questions by reconstructing the
scene and taking a microscopic examination of the system to determine why the phenomenon
occurs. You cannot accomplish this with a real system because you cannot see or control it in its
entirety.
Explore Possibilities. One of the greatest advantages of using simulation software is that once
you have developed a valid simulation model, you can explore new policies, operating
procedures, or methods without the expense and disruption of experimenting with the real
system. Modifications are incorporated in the model, and you observe the effects of those
changes on the computer rather than the real system.
Diagnose Problems. The modern factory floor or service organization is very complex. So
complex that it is impossible to consider all the interactions taking place in one given moment.
Simulation allows you to better understand the interactions among the variables that make up
such complex systems. Diagnosing problems and gaining insight into the importance of these
variables increases your understanding of their important effects on the performance of the
overall system.
The last three claims can be made for virtually all modeling activities, queuing, linear
programming, etc. However, with simulation the models can become very complex and, thus,
have a higher fidelity, i.e., they are valid representations of reality.
Identify Constraints. Production bottlenecks give manufacturers headaches. It is easy to forget
that bottlenecks are an effect rather than a cause. However, by using simulation to perform
bottleneck analysis, you can discover the cause of the delays in work-in-process, information,
materials, or other processes.
Develop Understanding. Many people operate with the philosophy that talking loudly, using
computerized layouts, and writing complex reports convinces others that a manufacturing or
service system design is valid. In many cases these designs are based on someone's thoughts
about the way the system operates rather than on analysis. Simulation studies aid in providing

understanding about how a system really operates rather than indicating an individual's
predictions about how a system will operate.
Visualize the Plan. Taking your designs beyond CAD drawings by using the animation features
offered by many simulation packages allows you to see your facility or organization actually
running. Depending on the software used, you may be able to view your operations from various
angles and levels of magnification, even 3-D. This allows you to detect design flaws that appear
credible when seen just on paper on in a 2-D CAD drawing.
Build Consensus. Using simulation to present design changes creates an objective opinion. You
avoid having inferences made when you approve or disapprove of designs because you simply
select the designs and modifications that provided the most desirable results, whether it be
increasing production or reducing the waiting time for service. In addition, it is much easier to
accept reliable simulation results, which have been modeled, tested, validated, and visually
represented, instead of one person's opinion of the results that will occur from a proposed design.
Prepare for Change. We all know that the future will bring change. Answering all of the "whatif" questions is useful for both designing new systems and redesigning existing systems.
Interacting with all those involved in a project during the problem-formulation stage gives you
an idea of the scenarios that are of interest. Then you construct the model so that it answers
questions pertaining to those scenarios. What if an automated guided vehicle (AGV) is removed
from service for an extended period of time? What if demand for service increases by 10
percent? What if....? The options are unlimited.
Wise Investment. The typical cost of a simulation study is substantially less than 1% of the total
amount being expended for the implementation of a design or redesign. Since the cost of a
change or modification to a system after installation is so great, simulation is a wise investment.
Train the Team. Simulation models can provide excellent training when designed for that
purpose. Used in this manner, the team provides decision inputs to the simulation model as it
progresses. The team, and individual members of the team, can learn by their mistakes, and learn
to operate better. This is much less expensive and less disruptive than on-the-job learning.
Specify Requirements. Simulation can be used to specify requirements for a system design. For
example, the specifications for a particular type of machine in a complex system to achieve a
desired goal may be unknown. By simulating different capabilities for the machine, the
requirements can be established.
B. Disadvantages
The disadvantages of simulation include the following:
Model Building Requires Special Training. It is an art that is learned over time and through
experience. Furthermore, if two models of the same system are constructed by two competent
individuals, they may have similarities, but it is highly unlikely that they will be the same.
Simulation Results May Be Difficult to Interpret. Since most simulation outputs are
essentially random variables (they are usually based on random inputs), it may be hard to
determine whether an observation is a result of system interrelationships or randomness.
Simulation Modeling and Analysis Can Be Time Consuming and Expensive.

Skimping on resources for modeling and analysis may result in a simulation model and/or
analysis that is not sufficient for the task.
Simulation May Be Used Inappropriately. Simulation is used in some cases when an analytical
solution is possible, or even preferable. This is particularly true in the simulation of some waiting
lines where closed-form queueing models are available, at least for long-run evaluation.
Pseudo-Random Number Generation
Probability is used to express our confidence in the outcome of some random event as a real
number between 0 and 1. An outcome that is impossible has probability 0; one that is inevitable
has probability 1. Sometimes, the probability of an outcome is calculated by recording the
outcome for a very large number of occurrences of that random event (the more the better), and
then taking the ratio of the number of events in which the given outcome occurred to the total
number of events.
It is also possible to determine the probability of an event in a non-experimental way, by listing
all possible non-overlapping outcomes for the experiment and then using some insight to assign a
probability to each outcome.
For example, we can show that the probability of getting heads twice during three coin tosses
should be 3/8 from the following argument. First, we list all eight possible outcomes for the
experiment (TTT, TTH, THT, THH, HTT, HTH, HHT, HHH). Next, we assign equal probability
to each outcome (i.e., 1/8), because a fair coin should come up heads half the time, and the
outcome of one coin toss should have no influence on another. (In general, when we have no
other information to go on, each possible outcome for an experiment is assigned the same
probability.)
And finally, we observe that the desired event (namely two Hs) includes three of these
outcomes, so that its
probability should be the sum of the probabilities for those outcomes, namely 3/8.
A random variable is a variable, say X, whose value, x, depends on the outcome of some
random event. For example, we can define a random variable that takes on the value of 1
whenever heads occurs and 0 otherwise.
Generating Uniform Pseudo-Random Numbers
Since the result of executing any computer program is in general both predictable and repeatable
the idea that a computer can be used to generate a sequence of random numbers seems to be a
contradiction. There is no contradiction, however, because the sequence of numbers that is
actually generated by a computer is entirely predictable once the algorithm is known. Such
algorithmically generated sequences are called pseudo-random sequences because they appear
to be random in the sense that a good pseudo-random number sequence can pass most statistical
tests designed to check that their distribution is the same as the intended distribution. On the
other hand, to call them random numbers is no worse than it is to label floating point numbers
as real numbers in a programming language.
It is worth noting that the availability of pseudo-random (rather than truly random) numbers is
actually an advantage in the design of simulation experiments, because it means that our
random numbers are reproducible.
Thus, to compare two strategies for operating our service station, we can run two experiments
with different numbers of pumps but with exactly the same sequence of customers. Were we to
attempt the same comparison using an actual service station, we would have to try the two
alternatives one after the other, with a different sequence of customers (e.g., those that came on
Tuesday instead of those that came on Monday), making it difficult to claim that a change in our

profits was due to the change in strategy rather than the differences in traffic patterns between
Mondays and Tuesdays.
In general, most pseudo-random number generators produce uniform values. (This does not lead
to a loss of generality because, as we shall see in the next section, uniform values can be
transformed into other distributions.)
Since only finite accuracy is possible on a real computer, we cannot generate continuous random
variables. However, it is possible to generate pseudo-random integers, xk , uniformly between 0
and some very large number (say m), which we use directly as a discrete uniform distribution, or
transform into the fraction xk/m to be used as an approximation to a uniform continuous
distribution between 0 and 1.
The most popular method of generating uniform pseudo-random integers is as follows. First, an
initial value x 0, the seed, is chosen. Thereafter, each number, xk , is used to generate the next
number in the sequence, xk +1, from the relation
xk +1 := (a * xk + c) mod m
The only difficult part is choosing the constants a, c and m. These must be chosen both to ensure
that the period of the random number generator (i.e., the number of distinct values before the
sequence repeats) is large, and also to ensure that consecutive numbers appear to be independent.
Note that the period can be at most m, because there are only m distinct values between 0 and
m1.
Progress in a simulation study
1. Problem formulation. Every simulation study begins with a statement of the problem. If the
statement is provided by those that have the problem (client), the simulation analyst must take
extreme care to insure that the problem is clearly understood. If a problem statement is prepared
by the simulation analyst, it is important that the client understand and agree with the
formulation. It is suggested that a set of assumptions be prepared by the simulation analyst and
agreed to by the client. Even with all of these precautions, it is possible that the problem will
need to be reformulated as the simulation study progresses.
2. Setting of objectives and overall project plan. Another way to state this step is "prepare a
proposal." This step should be accomplished regardless of location of the analyst and client, viz.,
as an external or internal consultant.
The objectives indicate the questions that are to be answered by the simulation study. The project
plan should include a statement of the various scenarios that will be investigated. The plans for
the study should be indicated in terms of time that will be required, personnel that will be used,
hardware and software requirements if the client wants to run the model and conduct the
analysis, stages in the investigation, output at each stage, cost of the study and billing
procedures, if any.
3. Model conceptualization. The real-world system under investigation is abstracted by a
conceptual model, a series of mathematical and logical relationships concerning the components
and the structure of the system. It is recommended that modeling begin simply and that the
model grow until a model of appropriate complexity has been developed. For example, consider
the model of a manufacturing and material handling system. The basic model with the arrivals,
queues and servers is constructed. Then, add the failures and shift schedules. Next, add the
material-handling capabilities. Finally, add the special features. Constructing an unduly complex
model will add to the cost of the study and the time for its completion without increasing the

quality of the output. Maintaining client involvement will enhance the quality of the resulting
model and increase the client's confidence in its use.
4. Data collection. Shortly after the proposal is "accepted" a schedule of data requirements
should be submitted to the client. In the best of circumstances, the client has been collecting the
kind of data needed in the format required, and can submit these data to the simulation analyst in
electronic format. Oftentimes, the client indicates that the required data are indeed available.
However, when the data are delivered they are found to be quite different than anticipated. For
example, in the simulation of an airline-reservation system, the simulation analyst was told "we
have every bit of data that you want over the last five years." When the study commenced, the
data delivered were the average "talk time" of the reservationist for each of the years. Individual
values were needed, not summary measures. Model building and data collection are shown as
contemporaneous in . This is to indicate that the simulation analyst can readily construct the
model while the data collection is progressing.
5. Model translation. The conceptual model constructed in Step 3 is coded into a computer
recognizable form, an operational model.
6. Verified? Verification concerns the operational model. Is it performing properly? Even with
small textbook sized models, it is quite possible that they have verification difficulties. These
models are orders of magnitude smaller than real models (say 50 lines of computer code versus
2,000 lines of computer code). It is highly advisable that verification take place as a continuing
process. It is ill advised for the simulation analyst to wait until the entire model is complete to
begin the verification process. Also, use of an interactive run controller, or debugger, is highly
encouraged as an aid to the verification process.
7. Validated? Validation is the determination that the conceptual model is an accurate
representation of the real system. Can the model be substituted for the real system for the
purposes of experimentation? If there is an existing system, call it the base system, then an ideal
way to validate the model is to compare its output to that of the base system. Unfortunately, there
is not always a base system. There are many methods for performing validation.
8. Experimental design. For each scenario that is to be simulated, decisions need to be made
concerning the length of the simulation run, the number of runs (also called replications), and the
manner of initialization, as required.
9. Production runs and analysis. Production runs, and their subsequent analysis, are used to
estimate measures of performance for the scenarios that are being simulated.
10. More runs? Based on the analysis of runs that have been completed, the simulation analyst
determines if additional runs are needed and if any additional scenarios need to be simulated.
11. Documentation and reporting. Documentation is necessary for numerous reasons. If the
simulation model is going to be used again by the same or different analysts, it may be necessary
to understand how the simulation model operates. This will enable confidence in the simulation
model so that the client can make decisions based on the analysis. Also, if the model is to be
modified, this can be greatly facilitated by adequate documentation. The result of all the analysis
should be reported clearly and concisely. This will enable the client to review the final
formulation, the alternatives that were addressed, the criterion by which the alternative systems
were compared, the results of the experiments, and analyst recommendations, if any.

12. Implementation. The simulation analyst acts as a reporter rather than an advocate. The
report prepared in step 11 stands on its merits, and is just additional information that the client
uses to make a decision. If the client has been involved throughout the study period, and the
simulation analyst has followed all of the steps rigorously, then the likelihood of a successful
implementation is increased.

Example : An inventory control system is a process for managing and locating objects or
materials. In common usage, the term may also refer to just the software components.
Modern inventory control systems often rely upon barcodes and radio-frequency identification
(RFID) tags to provide automatic identification of inventory objects. Inventory objects could
include any kind of physical asset: merchandise, consumables, fixed assets, circulating tools,
library books, or capital equipment. To record an inventory transaction, the system uses a
barcode scanner or RFID reader to automatically identify the inventory object, and then collects
additional information from the operators via fixed terminals (workstations), or mobile
computers.
The new trend in inventory management is to label inventory and assets with QR Code, and use
smartphones to keep track of inventory count and movement. These new systems are especially
useful for field service operations, where an employee needs to record inventory transaction or
look up inventory stock in the field, away from the computers and hand-held scanners.

Module-II
(12 Lectures)
Objective: To use the Probability concepts in Simulation
Events that cannot be predicted precisely are often called random. Many if not most of the inputs
to, and processes that occur in systems are to some extent random. Hence, so too are the outputs
or predicted impacts, and even peoples reactions to those outputs or impacts.
Probability Concepts and Methods

The basic concept in probability theory is that of the random variable. By definition, the value of
a random variable cannot be predicted with certainty. It depends, at least in part, on the outcome
of a chance event. Examples are: (1) the number of years until the flood stage of a river washes
away a small bridge; (2) the number of times during a reservoirs life that the level of the pool
will drop below a specified level; (3) the rainfall depth next month; and (4) next years maximum
flow at a gauge site on an unregulated stream. The values of all of these random events or
variables are not knowable before the event has occurred. Probability can be used to describe the
likelihood that these random variables will equal specific values or be within a given range of
specific values.
Distributions of Random Events
Given a set of observations to which a distribution is to be fit, one first selects a distribution
function to serve as a model of the distribution of the data. The choice of a distribution may be
based on experience with data of that type, some understanding of the mechanisms giving rise to
the data, and/or examination of the observations themselves. One can then estimate the
parameters of the chosen distribution and determine if the fitted distribution provides an
acceptable model of the data.
Stochastic Processes
Historical records of rainfall or stream flow at a particular site are a sequence of observations
called a time series. In a time series, the observations are ordered by time, and it is generally the
case that the observed value of the random variable at one time influences ones assessment of
the distribution of the random variable at later times. This means that the observations are not
independent. Time series are conceptualized as being a single observation of a stochastic process,
which is a generalization of the concept of a random variable.
Stochastic Variables, Discrete and Continuous Probability function:
The variables that can change with certain probability are called stochastic variables (random
variables).
In discrete system simulation we use the probability mass function. If a random variable is a
discrete variable, its probability distribution is called a discrete probability distribution.
An example will make this clear. Suppose you flip a coin two times. This simple statistical
experiment can have four possible outcomes: HH, HT, TH, and TT. Now, let the random variable
X represent the number of Heads that result from this experiment. The random variable X can
only take on the values 0, 1, or 2, so it is a discrete random variable.
The probability distribution for this statistical experiment appears below.
Number of heads
Probability
0
0.25
1
0.50
2
0.25
The above table represents a discrete probability distribution because it relates each value of a
discrete random variable with its probability of occurrence. Binomial and Poisson distribution
will also be covered here.
When variable being observed is continuous then an infinite number of possible values can be
assumed by the function and we describe the variable by a probability density function. On
integrating it within the range we get the cumulative distribution function.
Numerical evaluation of Continuous Probability function and continuous uniformly distributed
random number:

The purpose of generating a probability function into a simulation is to generate random numbers
with that particular distribution. Customary way of organizing data derived from observations is
to display them as a frequency distribution. Relative frequency distribution is a better approach.
By a continuous uniform distribution we mean that the probability of a variable, X, falling in any
interval within a certain range of values is proportional to the ratio of the interval size to the
range; that is, every point in the range is equally likely to be chosen.
Random number generators: Linear Congruential Generator, Mid Square Method, Multiplicative
Congruential Generator, Rejection Method:
Linear congruential generator (LCG) is an algorithm that yields a sequence of pseudorandomized numbers calculated with a discontinuous piecewise linear equation. The method
represents one of the oldest and best-known pseudorandom number generator algorithms. The
theory behind them is relatively easy to understand, and they are easily implemented and fast,
especially on computer hardware which can provide modulo arithmetic by storage-bit truncation.
The generator is defined by the recurrence relation:

where

is the sequence of pseudorandom values, and

the "modulus"
the "multiplier"
the "increment"
the "seed" or "start value"

are integer constants that specify the generator. If c = 0, the generator is often called
a multiplicative congruential generator (MCG), or Lehmer RNG. If c 0, the method is called
a mixed congruential generator.
Mid-square method is a method of generating pseudorandom numbers. In practice it is not a
good method, since its period is usually very short and it has some severe weaknesses, such as
the output sequence almost always converging to zero. Here we take the mid of a number as the
seed value and then square it to generate a random number and accordingly continue till we have
not obtained the desired set of random numbers.
The rejection sampling method generates sampling values from an arbitrary probability
distribution function
that

by using an instrumental distribution

where

, under the only restriction

is an appropriate bound on

Rejection sampling is usually used in cases where the form of

Instead

of

sampling

directly

from

the

distribution

makes sampling difficult.

,

we

use

an

envelope

distribution
where sampling is easier. These samples from
are
probabilistically accepted or rejected.
Testing of Random Numbers:
Chi-square is a statistical test commonly used to compare observed data with data we would
expect to obtain according to a specific hypothesis. For example, if, according to Mendel's laws,
you expected 10 of 20 offspring from a cross to be male and the actual observed number was 8

males, then you might want to know about the "goodness to fit" between the observed and
expected. Were the deviations (differences between observed and expected) the result of chance,
or were they due to other factors. How much deviation can occur before you, the investigator,
must conclude that something other than chance is at work, causing the observed to differ from
the expected. The chi-square test is always testing what scientists call the null hypothesis, which
states that there is no significant difference between the expected and observed result.
The formula for calculating chi-square ( x2) is:
x2= (o-e)2/e
That is, chi-square is the sum of the squared difference between observed (o) and the expected
(e) data (or the deviation, d), divided by the expected data in all possible categories. Chi-squared
distribution, showing X2 on the x-axis and P-value on the y-axis.
A chi-squared test, also referred to as test (or chi-square test), is any statistical hypothesis
test in which the sampling distribution of the test statistic is a chi-squared distribution when the
null hypothesis is true. Also considered a chi-squared test is a test in which this is asymptotically
true, meaning that the sampling distribution (if the null hypothesis is true) can be made to
approximate a chi-squared distribution as closely as desired by making the sample size large
enough. The chi-squared (I) test is used to determine whether there is a significant difference
between the expected frequencies and the observed frequencies in one or more categories. Does
the number of individuals or objects that fall in each category differ significantly from the
number you would expect? Is this difference between the expected and observed due to sampling
variation, or is it a real difference?

Exact chi-squared distribution

One case where the distribution of the test statistic is an exact chi-squared distribution is the test
that the variance of a normally distributed population has a given value based on a sample
variance. Such a test is uncommon in practice because values of variances to test against are
seldom known exactly.

Chi-squared test for variance in a normal population

If a sample of size n is taken from a population having a normal distribution, then there is a result
which allows a test to be made of whether the variance of the population has a pre-determined
value. For example, a manufacturing process might have been in stable condition for a long
period, allowing a value for the variance to be determined essentially without error. Suppose that
a variant of the process is being tested, giving rise to a small sample of n product items whose
variation is to be tested. The test statistic T in this instance could be set to be the sum of squares
about the sample mean, divided by the nominal value for the variance (i.e. the value to be tested
as holding). Then T has a chi-squared distribution with n 1 degrees of freedom. For example if
the sample size is 21, the acceptance region for T for a significance level of 5% is the interval
9.59 to 34.17.

Chi-squared test for independence and homogeneity in tables

Suppose a random sample of 650 of the 1 million residents of a city is taken, in which every
resident of each of four neighborhoods, A, B, C, and D, is equally likely to be chosen. A null
hypothesis says the randomly chosen person's neighborhood of residence is independent of the
person's occupational classification, which is either "blue collar", "white collar", or "service".
The data are tabulated:

Let us take the sample proportion living in neighborhood A, 150/650, to estimate what
proportion of the whole 1 million people live in neighborhood A. Similarly we take 349/650 to
estimate what proportion of the 1 million people are blue-collar workers. Then the null
hypothesis independence tells us that we should "expect" the number of blue-collar workers in
neighborhood A to be

Then in that "cell" of the table, we have

The sum of these quantities over all of the cells is the test statistic. Under the null hypothesis, it
has approximately a chi-squared distribution whose number of degrees of freedom is
If the test statistic is improbably large according to that chi-squared distribution, then one rejects
the null hypothesis of independence.
A related issue is a test of homogeneity. Suppose that instead of giving every resident of each of
the four neighborhoods an equal chance of inclusion in the sample, we decide in advance how
many residents of each neighborhood to include. Then each resident has the same chance of
being chosen as do all residents of the same neighborhood, but residents of different
neighborhoods would have different probabilities of being chosen if the four sample sizes are not
proportional to the populations of the four neighborhoods. In such a case, we would be testing
"homogeneity" rather than "independence". The question is whether the proportions of bluecollar, white-collar, and service workers in the four neighborhoods are the same. However, the
test is done in the same way.

The Kolmogorov-Smirnov Test

The Kolmogorov-Smirnov test is designed to test the hypothesis that a given data set could have
been drawn from a given distribution. Unlike the chi-square test, it is primarily intended for use
with continuous distributions and is independent of arbitrary computational choices such as bin
width.
Suppose that we had collected four data points and sorted them into increasing order to get the data
set {1.2, 3.1, 5.1, 6.7}. From the pattern of our data alone, we might guess that, if we continued to
collect data from this process, 0-25% of our observations would be less than or equal to 1.2, 2550% would be less than or equal to 3.1, etc.
Perhaps we would like to compare this empirical pattern to the pattern we would expect to observe
if the data points were drawn from a given theoretical distribution; say, an exponential distribution
with a mean of 5 (i.e., l = 1/5 = 0.2). If data points were drawn from this exponential distribution,

what fraction would we expect to see below 1.2? Below 3.1? These figures can be computed from
the cumulative distribution function for the exponential distribution:
F(1.2) = 1 - e-(0.2)(1.2) = 0.21
F(3.1) = 1 - e-(0.2)(3.1) = 0.46
F(5.1) = 1 - e-(0.2)(5.1) = 0.64
F(6.7) = 1 - e-(0.2)(6.7) = 0.74
We compare this to our empirical pattern in Figure 1. The first three rows contain the data points
and our highest and lowest estimates of the fraction of the data that would fall below each point.
The fourth row contains the result of plugging the data points into the theoretical distribution under
consideration (in this case, the exponential distribution with a mean of 5). These values are
the theoretical estimate of what fraction should fall below each data point. The fifth row is
obtained by comparing the fourth row to the second and third rows. Is 0.21 near 0? Near 0.25? We
take the absolute value of the larger of the two deviations. For example, in the first column, we get
|0 = 0.21| = 0.21
|0.25 - 0.21| = 0.04
so, the larger deviation is 0.21. This gives an idea of how far our empirical pattern is from our
theoretical pattern.
FIGURE 1: Computing D for the Kolmogorov-Smirnov test.
Row 1:

Data point x1

1.2

3.1

5.1

Row 2:

Empirical fraction falling below data

point (low estimate)

0.25

0.50

Row 3:

Empirical fraction falling below data

point (high estimate)

0.25

0.50

0.75

Row 4:

F(x1)

0.21

0.46

0.64

Row 5:

Largest deviation

0.21

0.21

0.14

Row 6:

Overall largest deviation (D)

Next, we look over the fifth row to find the largest overall deviation (D). The largest error, 0.26, is
the value of our test statistic. Is this measure of "error" large or small for this situation? To make
this judgment, we compare our computed value of this test statistic to a critical value from the
table in Appendix A. Setting a=0.1 and noting that our sample size is n=4, we get a critical value
of D4,1.0 = 0.565. Since our test statistic, D=0.26, is less than 0.565, we do not reject the hypothesis
that our data set was drawn from the exponential distribution with a mean of 5.
In general, we use the Kolmogorov-Smirnov test to compare a data set to a given theoretical
distribution by filling in a table as follows:
Row 1: Data set sorted into increasing order and denoted as xi, where i=1,...,n.

 Row 2: Smallest empirical estimate of fraction of points falling below xi, and computed as (i1)/n for i=1,...,n (e.g. if n=4, this row contains 0, 0.25, 0.50, and 0.75).
Row 3: Largest empirical estimate of fraction of points falling below xi and computed
as i/n for i=1,...,n (e.g., if n=4, this row contains 0.25, 0.50, 0.75, and 1.0).
Row 4: Theoretical estimate of fraction of points falling below xi and computed as F(xi),
where F(x) is the theoretical distribution function being tested.
Row 5: Absolute value of difference between row 2 and row 4 or between row 3 and row 4,
whichever is larger. This is a measure of "error" for this data point.
Row 6: The largest "error" from row 5, which gives the test statistic D.
Once this table has been completed, the test statistic D can be compared to the critical value from a
statistical table. If the test statistic is larger than the critical value, then we reject the hypothesis
that the data set was drawn from the given theoretical distribution; otherwise we do not reject the
hypothesis.*
In the preceding example, the parameter of the theoretical distribution (i.e., l = 1/5) was not
estimated from the data set. In most cases, the critical values in the Kolmogorov-Smirnov table are
only valid for testing distributions whose parameters have not been estimated from the data. Had
we, for example, used a maximum-likelihood estimate formula to compute l before testing the fit
of the distribution, we could not have used the Kolmogorov-Smirnov test. Modified versions of the
Kolmogorov-Smirnov test have been developed for testing the fit of a few theoretical distributions
in the case where parameter values have been estimated from the data. While we will not cover
these modified tests, the ideas behind them are similar to those we have discussed, and many
popular statistical software packages perform them.
Generation of Stochastic Variants:
For simulation experiments (including Monte Carlo) it is necessary to generate random numbers
(as values of variables). The problem is that the computer is highly deterministic machine basically, behind each process there is always an algorithm, deterministic computation changing
inputs to outputs, therefore it is not easy to generate uniformly spread random numbers over a
defined interval or set.
Random number generator is a device capable of producing a sequence of numbers which can not
be "easily" identified with deterministic properties. This sequence is then called Sequence of
stochastic numbers.
The algorithms typically rely on pseudo random numbers, computer generated numbers mimicking
true random numbers, to generate a realization, one possible outcome of a process.
Methods for obtaining random numbers exist for a long time and are used in many different fields
(like gaming). However, these numbers suffer from certain bias. Currently the best methods,
expected to produce truly random sequences are natural methods that take advantage of the
random nature of quantum phenomena.
Arrival Patterns Service Times:

Queuing models
Distribution of request inter-arrival times
The problem: Let us assume that we want to model a system consisting of many users and a server,
and we are interested in the response time of the server which is the time between the moment that
a user sends a request to the server to the moment when the server has completed its answer (we

ignore any communication delays).

Now, let us assume that the service time of the server is a constant (1 second), that is, it always
takes exactly 1 second to provide an answer to a request. This means the distribution of the service
time is a delta-function with a peak at t = 1 second.
If the users organize themselves in such a way that the next request will be sent to the server only
after the last answer was received, then the response time will always be equal to the service time.
However, if the users do not coordinate their requests, it may happen that a user sends a request
before the last answer was produced. This means that the server is busy when the request arrives
and the request has to wait (in some kind of input queue) before it will be processed. Therefore the
response time, in this case, will be larger than the service time (service time plus waiting time). We
see that the arrival pattern of the request has an impact on the average response time of the system.
Poisson inter-arrival pattern: It can be shown that, if there are very many users without any
coordination among them, the inter-arrival time (that is, the time between the arrival of one request
up to the arrival of the next request) has an exponential distribution of the form Pt = alpha * exp(alpha * t) where alpha is the arrival rate. In fact, the average of this inter-arrival distribution is
1/alpha which is the average time between two consecutive requests. We see that in this case there
is a good probability that the inter-arrival time is much smaller than the average. Therefore it
would be interesting to determine what the average waiting time of requests in the server input
queue would be. An answer is given by the queuing models.
Important conclusion: One cannot say what the response time of a server is without knowing the
inter-arrival pattern of the requests. The performance guarantee of the server depends on the
hypothesis about its environment (which determines the distribution of incoming requests).

Random variate
In the mathematical fields of probability and statistics, a random variate is a particular outcome
of a random variable: the random variates which are other outcomes of the same random variable
might have different values. Random variates are used when simulating processes driven by
random influences (stochastic processes). In modern applications, such simulations would derive
random variates corresponding to any given probability distribution from computer procedures
designed to create random variates corresponding to a uniform distribution, where these
procedures would actually provide values chosen from a uniform distribution of pseudorandom
numbers.
Procedures to generate random variates corresponding to a given distribution are known as
procedures for random variate generation or pseudo-random number sampling.
In probability theory, a random variable is a measurable function from a probability space to a
measurable space of values that the variable can take on. In that context, and in statistics, those
values are known as a random variates, or occasionally random deviates, and this represents a
wider meaning than just that associated with pseudorandom numbers.

Definition
Devroye defines a random variate generation algorithm (for real numbers) as follows:
Assume that
1.
Computers can manipulate real numbers.
2.
Computers have access to a source of random variates that are uniformly
distributed on the closed interval
.
Then a random variate generation algorithm is any program that halts almost surely and
exits with a real number X. This X is called a random variate.

(Both assumptions are violated in most real computers. Computers necessarily lack the ability to
manipulate real numbers, typically using floating point representations instead. Most computers
lack a source of true randomness (like certain hardware random number generators), and instead
use pseudorandom number sequences.)
The distinction between random variable and random variate is subtle and is not always made in
the literature. It is useful when one wants to distinguish between a random variable itself with an
associated probability distribution on the one hand, and random draws from that probability
distribution on the other, in particular when those draws are ultimately derived by floating-point
arithmetic from a pseudo-random sequence.

Random Variate Generation

It is assumed that a distribution is completely specified and we wish to generate
samples from this distribution as input to a simulation model.
Techniques
Inverse Transformation
Acceptance-Rejection
Convolution
All these techniques assume that a source of uniform (0, 1) random numbers is
available; R1, R2,..., where each Ri has:
1 , 0 x 1
pdf: fR(x) =
0 , otherwise

and

0 , x < 0
cdf: FR(x) = x , 0 x 1
1 , x > 1
Note: The random variable may be either discrete or continuous.
If the random variable is discrete, ==>
x take on a specific value, and F(x) is a step Fn
If F(x) is continuous over the domain x, ==>
f(x) = dF(x) / dx
and
the derivative f(x) is called the pdf.
Mathematically, the cdf is:
F(x) = P(X x) =
, where F(x) is defined over the range 0 F(x) 1, and f(t) represents the
value of the pdf of the variable x, when X = t.
Example #1
Generate random variates x with density function f(x) = 2x, 0 x 1

Solution:
F(x) =
= x2 , 0 x 1
Now set F(x) = R ==>
R = x2
-1
Next, solve for x, ==> x = F (R) = R, 0 r 1
\Values of x with pdf f(x) = 2x can be generated by taking the square root of the random, R.
Example #2
Generate random variates x with density function
f(x) = e-x , 0 x
0,
x<0
Solution:
F(x)

f(x) =

=
= 1 - e-x ,
0x
=0
x < 0Now set F(x) = R
Next solve for x, ==>
1 - e-x = R
e-x = 1 - R
- x = ln(1 - R)
x = - {ln(1 - R)} /
or
= - {ln(R)} /
Uniform Distribution
Consider a random variable X that is uniformly distributed on the interval [a, b]
pdf: f(x) = 1/ (b-a)
ax b
0,
otherwise
To generate random variates:
Step 1.
F(x) =
0
x< a
(x - a) / (b - a) , a x b
1
x>b
Step 2.
F(x) = (x - a) / (b - a) = R
Step 3.
X = a + (b - a) R

Module III
Objective: Know about simulation Language like GPSS:
1. Discrete events
A model used in discrete system simulation has a number of numbers to represent the state of a
system. A number used to represent some aspect of the system state is called a state descriptor.
Some state descriptors range over values that have physical significance, such as the number
representing the count of documents .As the simulation proceeds ,the state descriptors change
value. A discrete event is a set of circumstances, that causes an instantaneous change in one or
more system state descriptors. It is possible that two different events occur simultaneously, or are
modeled as being simultaneous, so that not all changes of state descriptors occurring
simultaneously necessarily belong to a single event.
2. Representation of time
The passage of time is recorded by a number referred to as a clock time.It is usually set to zero at
the beginning of a simulation and subsequently indicates how many units of simulated time have
passed since the beginning of the simulation. Here the simulation time means the clock time.
Ratio of simulated time to the real time taken can vary enormously.
Two basic methods exist for updating clock time. One method is to advance the clock to the time
at which the next event is due to occur. The other method is to advance the clock by small
intervals of time and determine at each interval whether an event is due to occur at that time.
Here the first method is called event oriented and the second method is called interval oriented.
Discrete system simulation is usually carried out with the use of event oriented method whereas
continuous system simulation uses interval oriented method. There is no firm rule to represent
time in simulation
3. Generation of arrival pattern
It is a important aspect of discrete system simulation. It is possible that an exact sequence of
arrivals has been specified for the simulation. Also the sequence of inputs may have been
generated from observations on a system. Computer system designs and especially the
programming components of the system are often tested with a record, gathered from a running
system that is representative of the sequence of operations the computer system will have to
execute. This approach is called trace driven simulation.
When there is no interaction between the exogenous arrivals and the endogenous events of the
system,a sequence of arrivals in preparation for the simulation is created. Usually simulation
proceeds by creating new arrivals as they are needed.
The exogenous arrival of an entity is defined as an event and the arrival time of the next entity is
recorded as one of the event times. When the clock time reaches this event time, the event of
entering the entity is calculated at once from the inter arrival time distribution.
4. Simulation of a telephone System
A telephony system has a number of telephones, connected to a switchboard by lines. Here the
object of the simulation will be to process a given number of calls and determine what proportion
are successfully completed, blocked, or found to be busy calls.
5. Delayed calls
Suppose we assume that calls that cannot be connected are not lost. Instead, they wait until they
can be connected. This case does not happen in a normal telephony system involving human

beings talking to each other. But it can happen to messages in a switching system that has store
and forward capability.
6. Priorities and parameters
A transaction has no particular identity. Each is treated by a block in the same manner as any
other transaction. Each transaction has one of 128 levels of priority, indicated by the numbers 0
to 127,with 0 being the lowest priority. At any point in the block diagram, the priority can be set
up or down to any of the levels by the PRIORITY block. The block is coded by putting the
priority in field A of the block.
It is also possible to designate the priority at the time a transaction is created by putting the
priority in the E field of the GENERATE block creating the transaction. If the field is left blank,
the priority is set to 0.
A transaction has also parameters, which carry numerical data that can affect the way the
transaction is processed by a block. The values are identified by the notation Pxn, where n is the
parameter number, and x is the type. If no declaration is made, the transaction have 12 half word
parameters.
All parameter values are zero at the time a transaction is created. A value is given to a parameter
when a transaction enters an ASSIGN block. The number of the parameter is given in the field A
of the ASSIGN block. The value of the parameter to be taken is given in the field B.
7. Standard Numerical Attributes
Every entity in a system has some attributes, for e.g. number of transactions in a storage or the
length of a queue, which are made available to the program user. These attributes, collectively
are called standard numerical attributes. Each type of SNA is identified by a one or two letter
code and a number.
Many uses can be made of SNAs. They provide the inputs to functions, and hence allow a great
variety of functional relationships to be introduced into the model. Values of the SNAs change as
the simulation proceeds. The program does not continuously maintain current value; it completes
value of SNAs at the time they are needed. Some example of GPSS SNA are C1, CHn, Fn, Kn,
M1 etc.
8. General description of GPSS
The system to be simulated in GPSS is described as a block diagram in which the blocks
represents the activities, and lines joining the blocks indicates the sequences in which the
activities can be executed.
The approach taken in GPSS is to define a set of 48 specific block types, each of which
represents a characteristic action of systems. The program user must draw a block diagram of the
system using only these block types.
Moving through the system being simulated that depends upon the nature of system. The
sequence of events in real time is reflected in movement of transaction s from block to block in
Fimulated time.
Transaction are created at one more GENERATE blocks and are removed from the
simulation at TERMINATE blocks . There can be many transactions simultaneously moving
through the block diagram. Each transaction is always positioned at a block and most blocks can
hold many transactions simultaneously. The transfer of transaction from one block to another
occurs instantaneously at a specific time or when some change of system condition occurs.
9. Facilities and Storages
A facility is defined as an entity that can be engaged by a single transaction at a time. A
storage is defined as an entity that can be occupied by many transaction at a time, up to some
predetermined limits. A transaction controlling a facility, however can be interrupted or
preempted by another transaction. In addition, both facilities and storages can be made
unavailable again, as occurs when a repairs has made.

There can be many instances of each type of entity to a limit set by the program
(usually 300). Individual entities are identified by number, a separate number sequence being
used for each type.
Four block types, SEIZE, RELEASE, ENTER, and LEAVE, concerned with using facilities and
storages. Field A in each case indicates which facility and storage intended, and choice is usually
marked in the flag attached to the system of blocks. The SEIZE block allows a transaction to
engage a facility if it is available. The RELEASE block allows the transaction to disengage the
facility. In an analogous manner, an ENTER block allows a transaction to occupy space in
storage, if it is available, and LEAVE block allows it to give up the space.
10. Gathering Statistics
Certain block types in GPSS are constructed for the purpose of gathering statistics about
the system performance, rather then of representing system actions. The QUEUE, DEPART,
MARK, and TABULATE blocks shown in fig 9-1 serve this purpose. They introduce two other
entities of GPSS program, queues and tables.
The QUEUE block increases and change is a unit change ; otherwise the value of field B
(>=1) is used. The program measured the average and maximum queue lengths and, if required,
the distribution of time spent on queue.
It is also described to measure the length of time taken by transaction to move through
the system or parts of the system, and this can be done with MARK and TABULATE blocks.
Each of these blocks types notes the time a transaction arrives at the block. The MARK block
simply notes the time of arrival on the transaction. The TABULATE block subtracts the time
noted by a MARK block from the time of arrival at the TABULATE block.

Module-IV
Objective: To focus on the simulation languages, analysis and testing of simulation outputs
and recent trends .
Simulation languages:
A simulation programming language (SPL) is a POL( Problem oriented language) with special
features. Simulation being a problem-solving activity with its own needs, programming
languages have been written to rake special features available. General Purpose Simulation
System (GPSS) (originally Gordon's Programmable Simulation System after creator Geoffrey
Gordon; the name was changed when it was decided to release it as a product) is a discrete time
simulation general-purpose programming language, where a simulation clock advances in
discrete steps. A system is modelled as transactions enter the system and are passed from one
service (represented by blocs) to another. This is particularly well-suited for problems such as a
factory. GPSS is less flexible than simulation languages such as Simula and SIMSCRIPT II.5 but
it is easier to use and more popular.
The GPSS world view
The GPSS code may surprise many programmers with experience in procedural, object oriented
or functional programming. The world is rather simulated with entities moving through the
model. These entities, called Transactions, are envisioned as moving from Block to Block, where
a Block is a line of code and represents unit actions that affects the Transaction itself or other
entities.
These other entities can be broadly classified in Resources, Computational entities and Statistical
entities.Resources, like Facilities and Storages represent limited capacity resources.
Computational entities, like Ampervariables (variables), Functions and random generators are
used to represent the state of Transactions or elements of their environment. Statistical entities,
like Queues or Tables (histograms) collect statistical information of interest.
Sample code
The following example, taken from Simulation using GPSS, is the "Hello world!" of GPSS and
will illustrate the main concepts.
The aim is to simulate one day of operation of a barber shop. Customers arrive in a random
constant flow, enter the shop, queue if the barber is busy, get their hair cut on a first-come firstserved basis, and then leave the shop. We wish to know the average and maximum waiting line,
as well as the number of customers.
SIMULATE
; Define model
*
* Model segment 1
*
GENERATE 18,6
; Customer arrive every 186 mn

QUEUE
Chairs
SEIZE
Joe
DEPART
Chairs
ADVANCE 16,4
RELEASE Joe
TERMINATE
*
*
*

Enter the line

Capture the barber
Leave the line
Get a hair cut in 164 mn
Free the barber
Leave the shop

Model segment 2
GENERATE 480
TERMINATE 1

*
*
*

;
;
;
;
;
;

; Timer arrives at time = 480 mn

; Shut off the run

Control cards
START
END

; Start one run

; End model

The "program" is comprised between the SIMULATE and END statements, and is divided into
"model segments" and "control cards".
The first segment models customers. The GENERATE block creates a flow of Transactions and
schedules them to enter the model with an inter-arrival time uniformly distributed over the range
186. It is the programmer's responsibility to interpret these transaction as customers and to
understand that the time is to be counted in minutes. The Transactions start their existence in the
GENERATE block and progress from Block to Block, according to certain rules, until they reach
a TERMINATE which remove them from the model.
Normally transactions progress from one block to the next one, so the customer transactions will
leave the GENERATE block to enter the QUEUE Chairs block. This block simulates a waiting
line, and collects statistics accordingly. In the example, it materialize a line of chairs and, at the
end of the simulation, we will know, among other things, the maximum queue size (how many
chairs are needed) and the average waiting time. The QUEUE block requires the name of the
queue as a parameter, because more than one queue may exist in the model. Each one is
associated with a DEPART block, which is triggered when the transaction leaves the queue.
GPSS remembers which transactions are in the queue, so that it possible to know the average
time spent, and to check that no buggy transaction is leaving a queue without previously entering
in it.
After the QUEUE chairs block, the transaction will try to proceed to the SEIZE Joe block,
a block simulating the capture of the Facility named Joe. Facilities model single servers of
capacity one. If the facility is busy, the SEIZE will deny the attempting transaction the right to
enter. In the example, the customer will wait in the QUEUE block. If it is free, or as soon as it
becomes available, the transaction will be allowed to capture the facility, mark it as busy to
others transactions and start to count the service time and other statistics, until the same
transaction passes the corresponding RELEASE Joe block.
The SEIZE / RELEASE pairs are linked by the facility name, because many independent
facilities may exist in the model. They can model operators, like a barber, a repairman, an agent,
but also pieces of equipment, like a crane, a gas station, an authorization document, etc., in fact
anything with capacity one. To simulate multiple parallel servers, like a team of five barbers, or
an oven with a capacity of 10, GPSS uses entities named STORAGEs.

After a customer seizes Joe, she proceeds to the next statement which is ADVANCE 16,4,
whose task is to freeze the entity for a prescribed length of time, here a random number picked
between 16-4=12 and 16+4=20mn. Other service time distributions are available through GPSS
FUNCTION (a somehow different notion than function in other programming languages). During
that time, other transactions will be allowed to move through the model, blocking some other
facilities that may exist in the model, but not Joe because this facility is busy with the frozen
customer. After the prescribed time, the customer will wake up, proceed to the next statement,
which will free Joe, and TERMINATE.
Then the next transaction on the previous block, that is a customer sitting on a chair, will be able
to SEIZE Joe. To select the "next" transaction, GPSS uses the first-come first-served basis,
with priority. Other selection policies can be programmed by direct manipulation of the future
event chain entity.
In parallel to this first segment, simulating the customer behavior, a second model segment
simulates the end of the day. At time 480mn = 8h a entity is GENERATEd, which will
TERMINATE on the next block. This time, the TERMINATE as a parameter of 1, meaning a
special counter is decreased by 1. When that counter reaches 0, the program stops and the output
is printed. This special counter is setup with the START statement. In the example, it is set to
one, thus the simulation will finish after one run of 480 mn in simulated time.
It indicates that Joe was busy 86.0% of the time, gave a hair cut to 26 customers and that hair cut
took 15.88 minutes on the average. Incidentally, Joe was cutting the hair of customer number 26
when the simulation was closed. No programming provisions were taken for the barber to finish
the hair cut before to close the shop.
It indicates also that a maximum of 1 customer was observed waiting his turn, in facts the
number of waiting customer was on the average 0.160. A total of 27 customers did enter the
queue, so that customer number 27 was still sitting, waiting his turn, when Joe closed the shop.
Out of these 27 customers, 12 were served without having to wait. In facts, the queue was empty
44.4% of the time. The average waiting time was 2.851 mn, and the average waiting time for the
15=27-12 customers who did really wait was 5.133 m
Output Analysis in Stochastic Simulation
Once the simulation program has been completely written and carefully tested, the actual
simulation experiment must be carried out. The idea of the experiment is to conduct one or more
simulation runs from which measurements are collected. The output from these runs is used to
assess the performance of the system (or sys-tems, if alternative strategies are being evaluated).
For example, one run from our service station simulator might have produced the following
output:
This
simulat
ion run
uses
three
pumps
and the
followi
ng
random
number
seeds:
1
2
3
4
Curre
Averag
Averag
nt
Total NoQueue Car >Car e
Number e
Pump Total

Lost

__________________________________________________________________________________________
__________________________________TimeCarsFractionTimeLitresBalkedWaitUsageProfi
tProfit
20000 382 0.529 52.356 34.342 67
54.877 0.875 221.34 46.62
40000 771 0.470 51.881 34.454 136
59.949 0.887 505.85 98.25
150.8
60000 1157 0.475 51.858 34.767 206
61.015 0.888 794.76 7
1088.2 213.5
80000 1554 0.476 51.480 35.052 290
60.764 0.884 6
1
10000
1376.5 283.3
0
1967 0.475 50.839 34.974 379
60.524 0.884 1
5
12000
1652.2 342.3
0
2351 0.477 51.042 34.957 451
60.753 0.882 5
8
14000
1934.7 408.0
0
2760 0.473 50.725 34.824 539
61.889 0.883 4
9
16000
2225.0 467.8
0
3175 0.468 50.394 34.682 625
62.332 0.886 5
6
18000
2524.1 521.0
0
3574 0.464 50.364 34.753 696
62.592 0.888 6
5
20000
2807.6 569.6
0
3957 0.468 50.543 34.746 769
61.654 0.887 8
1
2.9.1. Initial and Final Conditions
When calculating performance figures from a simulation run, care must be
taken to ensure that the results obtained are not systematically distorted because
of either the initial state of the system at the beginning of each simulation run or
the final state of the system when end-of-simulation is encountered.
For example, suppose we were concerned with the time for an auto to fill up
its tank during a busy time of day, such as afternoon rush hour before a holiday. If
we ran the simulation for a long time, we might find that under those operating
conditions an average of 15 autos are in the station having their tanks filled or
waiting for a free pump. But if the initial state for the simulation did not have any
autos in the service station, then it would be obvi-ous that the first few autos spent
less time waiting than a typical auto would. Thus, our estimate for the mean
system time would turn out too low unless we
i)

ran our simulation for such a long time that the measurements from
these first few customers had very little influence on the mean,

ii)

began our measurements part way through the simulation run, by

which point the system should have reached a state that is more
normal, or

iii) used the average values from one simulation run as the initial
conditions for the next simulation run so that, hopefully, the system is
in a normal state throughout the entire run.
In the output from our service station simulator, the effect of the initial
conditions (i.e., a completely idle sta-tion) are quite evident. Notice that even
though the first progress report was not printed until almost 400 cars had passed
through our service station, we still see a downward trend in the fraction of time
that the queue of waiting cars is empty, and an upward trend in both the average
waiting time and the utilization of the pumps.
The end conditions of the simulation must also be checked before we can
trust our calculated performance statistics. For example, the waiting times and

service times for customers in the system when end-of-simulation is reached may
or may not be counted, depending on the method used for recording these
measurements. If such cus-tomers are not counted, then we risk underestimating
the utilization of the server, because some customers who might have been part
way through their service time are ignored. If such customers are counted, then
we risk underestimating the mean waiting time and overestimating the mean
service rate because, in effect, we may be claiming that these customers were
served during our simulation run when in fact their service would not have been
completed until later.
In the case of our service station simulator, notice that we are adding the
service times for each car to the total service performed in the startService
procedure. Thus, we are always overestimating the pump utilization: in effect, we
are (optimistically) claiming that the entire service times for all cars not still in
the waiting queue have been completed by the end of simulation even those
cars still in service. If the run length is short enough, this choice of stopping
condition can lead to impossibly good results. For example, notice that in the
following data (representing a short run with a heavily loaded service station) the
pumps appear to be busy more than 100 percent of the time at the first few
progress reports!
This
simulat
ion run
uses
two
pumps
and the
followi
ng
random
number
seeds:
1
2
3
4
Car>Ca
Curre Tota
Averag
Averag
nt
l
NoQueue r
e
Number e
Pump Total Lost
! Time___ _ _Cars__ _ _ Fraction_____ _ _ _Time__ _ _ _ Litres____ _ _Balked____ _
_Wait__ _ _ _Usage___ _ Profit____ _ _Profit___ _ _ _ _ _ _ _ _ _
182.89
2000 39
0.072
51.282 31.508 11
7
1.002 -18.43 9.15
184.42
4000 82
0.039
48.780 31.037 28
9
1.004 2.70 20.93
190.62
6000 124
0.059
48.387 33.032 46
0
1.014 26.29 36.10
190.95
8000 162
0.068
49.383 34.455 62
9
1.005 50.97 48.57
205.27
10000 207
0.069
48.309 33.815 81
7
1.002 71.16 63.84
204.17
12000 240
0.067
50.000 34.090 92
6
1.005 92.20 72.33
215.16
14000 284
0.058
49.296 34.094 111
1
1.005 115.38 86.68
210.37
16000 304
0.088
52.632 34.567 113
7
0.993 134.53 88.18
210.08
18000 343
0.094
52.478 34.308 127
6
0.996 155.13 99.06
211.13
20000 382
0.084
52.356 34.227 140
8
0.997 178.59 108.28

Of course, we could easily replace the overestimate with an underestimate

by moving the accumulation of service time to the departure routine. However,
calculating the correct value for pump utilization is much trickier.
Another possibility that must be considered in assessing the end conditions
of the simulation is whether or not the system being simulated is overloaded.
Since only a finite period of simulated time elapses during each simula-tion run,
the average waiting time will always be finite, even if the system is in fact so
overloaded that the customer waiting times are growing without bound as
simulated time advances. Thus, our performance statistics are unreli-able if a
significant proportion of the customers are still in the system at the end of
simulation: either the simulation run was too short (and the effects of the initial
and final conditions are large), or the system was overload.
Confidence Intervals on the Mean of a Random Variable
Once we have carried out one or more simulation runs, and convinced
ourselves that the boundary conditions due to initial and final conditions have not
spoiled our results, we are left with the task of trying to conclude some-thing
about the values of various random variables in the model, such as the mean
waiting time for an auto in our service station, for example. It is simple to
calculate the sample mean
i.e., the average of the N measurements that we made of the value of the random
variable X. Similarly, we can calculate the sample variance:

by averaging the squared differences between each measurement and the sample
mean. But as we scan down the columns of measurements in the output, it is
obvious that these sample quantities vary as a function of (simulated) time.
Furthermore, if we repeat the experiment using different random number seeds,
we soon see that the measure-ments also vary from one run to the next.
Consequently, we should be skeptical of concluding that any of these sample
means or variances are equal to the true mean and variance of X that we set out to
find in the first place. And indeed to proceed further, we need to make some
assumptions about the distribution of X.

As a first step, suppose a helpful genie were to tell us that the random
variable X has a Normal distribution (i.e., a bell-shaped probability density
function). In this case, knowledge of the Normal distribution would allow us to
construct confidence intervals about the sample mean, say X* C, that contain
the true mean, say M, a specified fraction of the time (the greater the fraction, the
wider the resulting confidence interval about X+). This construction is based on
the following two facts about the Normal distribution with mean M and variance

2 . First, the proportion of measurements that fall within a distance of K from

M is known. For example, if we make a large number of measurements x 1 , x
2 , . . . , then approximately 95 percent of them will fall within the interval M
2. The key observation for us is that, since the distance from M to a
measurement xj is the same as the distance from x j to M, we can turn this result
around to say that 95 percent of the time M falls within the interval x j 2. A
table of results in this form is shown below in Figure 5. Second, the average of N
measurements taken from a Normal distribution with mean M and variance 2
also has a Normal distribution, but with mean M and variance 2 /N. Thus, if the
confidence interval ends up too wide, it can be tightened by averaging in
additional measurements. For example, using four times as many measurements
cuts the width of the confidence interval in half.
Figure 5. Intervals about a measurement x that will contain the true mean a given
fraction of the time, when the ran-dom variable has a Normal distribution with
mean M and variance 2
The problem with the result above is that even if we knew that the random
variable had a Normal distribution, we still couldnt write down a confidence
interval for M without knowing 2 (which depends on M). To get out of this
circular chain of reasoning, we need a method of estimating 2 that is based on
X0 instead of M. Clearly the sample variance, defined above, is not a very good
estimate, especially when the number of measurements is small. Indeed, if only
one measurement is available, then the sample variance is always zero, indicating
that we have not observed any variability in our measurements. But 2 = 0 means
something quite different, namely that X is not a random variable at all! The
solution is to use an unbiased estimate of the sample variance to represent 2 ,
i.e.,
N 23
N_1_1_j=11 xj X4562 ,
and then simply go ahead and use the results described above in Table 1. (By
dividing the sum by N 1 instead of N, it is clear that we cannot estimate 2 from
a single measurement.)
A more serious problem with the method outlined above is that the kinds of
random variables that usually come up in simulation models, such as the mean
time in system for autos at our service station, almost never have a Normal
distribution! Fortunately, the Central Limit Theorem in statistics tells us that
under fairly general cir-cumstances, the average of a large number of
measurements does have a Normal distribution even if the individual
measurements themselves do not have a Normal distribution. It follows that we
can still find a confidence interval if we carry out K independent repetitions of the
same simulation experiment and treat the K sample means X7j , j =1, . . . , K, as
the measurements.
For example, suppose we ran four more runs of the service station
simulation with 3 pumps, giving us, in total, five independent measurements of
the mean waiting time. The last line of output from each of these runs is
reproduced in the following table:
2

Current
NoQueue
Car>Car
Lost

Average

11

Number

Average

Total
Pump Total

8_Time8_8_8_8_8_8Cars_8_8_8_8Fraction_8_8_8_8_8_8_8_Time8_8_8_8_8_8Lit
res_8_8_8_8_8_Balked8_8_8_8_8_8_Wait8_8_8_8_8_Usage8_8_8_8_8Profit_8_8
_8_8_8_Profit8_8_8_8_8_8_8_8_8_8_8_8_
20000
2807.6 569.6
0
3957 0.468 50.543 34.746 769
61.654 0.887 8
1
20000
2870.2 662.0
0
4096 0.397 48.828 35.081 881
76.382 0.903 0
9
20000
2826.6 652.2
0
4062 0.432 49.237 34.850 848
66.428 0.899 9
9
20000
2821.6 594.9
0
3961 0.440 50.492 35.108 787
69.013 0.894 6
3
20000
2799.5 604.6
0
3972 0.429 50.352 34.886 790
66.098 0.890 3
2
Variance Reduction Techniques
In a simulation study, we are interested in one or more performance measures for some stochastic
model.Methods that one can use to reduce the variance of the estimator W . We will successively
describe the following techniques:

Common Random Numbers

Antithetic Variables
Control Variates
Conditioning
The first method is typically used in a simulation study in which we want to compare performance measures of two different systems. All other methods are also useful in the case that we
want to simulate a performance measure of only a single system.
Common Random Numbers
The idea of the method of common random numbers is that if we compare two different systems
with some random components it is in general better to evaluate both systems with the same realizations of the random components. Key idea in the method is that if X and Y are two random
variables, then
Var.X Y/ D Var.X/ C Var.Y/ 2Cov.X; Y/:
Hence, in the case that X and Y are positively correlated, i.e. Cov.X; Y/ > 0, the variance of X Y
will be smaller than in the case that X and Y are independent. In general, the use of common
random numbers leads to positive correlation of the outcomes of a simulation of two systems. As
a consequence, it is better to use common random numbers instead of independent random
numbers when we compare two different systems.
Let us illustrate the method using the following scheduling example. Suppose that a finite
number of N jobs has to be processed on two identical machines. The processing times of the
jobs are random variables with some common distribution function F . We want to compare the
completion time of the last job, Cmax, under two different policies. In the LPTF policy, we
always choose the remaining job with the longest processing time first, in the SPTF policy we
choose the remaining job with the shortest processing time first.

In Table 5 and Table 6 we compare for N D 10 and F .x / D 1 e x the estimators and confidence intervals for E .CmaxSPTF CmaxLPTF/ when we do not, resp. do, use common random
numbers.
We conclude that in this example the use of common random numbers reduces the standard
deviation of the estimator and hence also the length of the confidence interval with a factor 5.
30
# of runs
1000
10000
100000
1000000

mean
0.8138
0.8293
0.8487
0.8398

st. dev.
2.5745
2.4976
2.4990
2.4951

95% conf. int.

[0.645, 0.973]
[0.780, 0.878]
[0.833, 0.864]
[0.835, 0.845]

Table 5: Estimation of E .CmaxSPTF CmaxLPTF/ without using common random

numbers
# of runs
1000
10000
100000
1000000

mean
0.8559
0.8415
0.8394
0.8391

st. dev.
0.5416
0.5230
0.5164
0.5168

95% conf. int.

[0.822, 0.889]
[0.831, 0.852]
[0.836, 0.843]
[0.838, 0.840]

Table 6: Estimation of E .CmaxSPTF CmaxLPTF/ using common random numbers

When we want to use common random numbers, the problem of synchronization can arise:
How can we achieve that the same random numbers are used for the generation of the same
random variables in the two systems?
In the previous example, this synchronization problem did not arise. However, to illustrate
this problem, consider the following situation. In a G=G=1 queueing system the server can work
at two different speeds, v1 and v2. Aim of the simulation is to obtain an estimator for the
difference of the waiting times in the two situations. We want to use the same realizations of the
interarrival times and the sizes of the service requests in both systems (the service time is then
given by the sizes of the service request divided by the speed of the server). If we use the
program of the discrete event simulation of Section 3 of the G=G=1 queue, then we get the
synchronization problem because the order in which departure and arrival events take place
depends on the speed of the server. Hence, also the order in which interarrival times and sizes of
service requests are generated depend on the speed of the server.
The synchronization problem can be solved by one of the following two approaches:
% Use separate random number streams for the different sequences of random variables
needed in the simulation.
% Assure that the random variables are generated in exactly the same order in the two systems.
For the example of the G=G=1 queue, the first approach can be realized by using a separate
random number stream for the interarrival times and for the service requests. The second
approach can be realized by generating the service request of a customer already at the arrival

instant of the customer.

7.2 Antithetic Variables

The method of antithetic variables makes use of the fact that if U is uniformly distributed on .0;
1/ then so is 1 U and furthermore U and 1 U are negatively correlated. The key idea is that, if
31

W1 and W2 are the outcomes of two successive simulation runs, then

W 1 CW 2

1
V a r .W /

Cov.W; W / :
2 D 4
2 U
1 C 4
2 C 2
1
Hence, in the case that W1 and W2 are negatively correlated the variance of .W1 C W2/=2 will be
smaller than in the case that W1 and W2 are independent.
The question remains how we can achieve that the outcome of two successive simulation
runs will be negatively correlated. From the fact that U and 1 U are negatively correlated, we
may expect that, if we use the random variables U1; : : : ; Um to compute W1, the outcome of the
first simulation run, and after that 1 U1; : : : ; 1 Um to compute W2, the outcome of the
second simulation run, then also W1 and W2 are negatively correlated. Intuition here is that, e.g.,
in the simulation of the G=G=1 queue large realizations of the Ui 's corresponding to large
service times lead to large waiting times in the first run. Using the antithetic variables, this gives
small realizations of the 1 Ui 's corresponding to small service times and hence leading to small
waiting times in the second run.
We illustrate the method of antithetic variables using the scheduling example of the previous
subsection.
Var

# of runs
1000
10000
100000
1000000

mean
5.0457
5.0400
5.0487
5.0559

st. dev.
1.6201
1.6020
1.5997
1.5980

V a r .W /

95% conf. int.

[4.945, 5.146]
[5.009, 5.071]
[5.039, 5.059]
[5.053, 5.059]

Table 7: Estimation of E .CmaxLPTF/ without using antithetic variables

# of pairs
500
5000
50000
500000

mean
5.0711
5.04 9 7
5.05 4 6
5.05 4 6

st. dev. 95% conf. int.

0.7216 [5.008, 5.134]
0.6916 [5.030, 5.069]
0.6858 [5.049, 5.061]
0.6844 [5.053, 5.056]

Table 8: Estimation of E .CmaxLPTF/ using antithetic variables

In Table 7 and Table 8 we compare, again for N D 10 and F .x / D 1 e x , the estimators and
confidence intervals for E .CmaxLPTF/ when we do not, resp. do, use antithetic variables. So,
for example, we compare the results for 1000 independent runs with the results for 1000 runs
consisting of 500 pairs of 2 runs where the second run of each pair uses antithetic variables. We
conclude that in this example the use of antithetic variables reduces the length of the confidence
interval with a factor 1.5.
Finally, remark that, like in the method of common random numbers, the synchronization
problem can arise. Furthermore, it should be noted that the method is easier to implement if all
random variables are generated using the inversion transform technique (only one uniform
random number is needed for the realization of one random variable) than if we use, e.g., the
rejection method to generate random variables (a random number of uniform random numbers
are needed for the realization of one random number).

Control Variates
The method of control variates is based on the following idea. Suppose that we want to estimate
some unknown performance measure E . X / by doing K independent simulation runs, the i -th
A
n
un
bi
as
ra
ed
n
es
on d
o
ti
e
m
m
yi va
el ri
E at
di ab
E
. or
.
X fo
ng le
th X i w i X / D / r
i
.
e . K th
XD
i
ou
tp
D
ut
1
X
i

/
=
K
.
H
o
w
e
v
e
r
,
a
s
s
u
m
e
t
h
a
t
a
t
t
h
e

s
a
m
e
t
i
m
e
w
e
a
r
e
a
b
l
e
t
o
E . X / is
given by

a related
output
variable Y

si
m
ul
at
e

E .Y /

E .Y /

E .Y /

,
wi
it h

D1 Yi

is
kn
o
w
n.
If
w
e
de
no
te
by

,
w
h
er
e

/=K ,
then,
for
any
cons
tant
c,
the
quan
tity

is
also
an
unbi
ased

estimator of E . X /.
Furthermore, from the formula

c2

P
.X

ar

c.Y

E .Y ///

.X/

Y/

Var

Va

C
2

. X ; Y /:

C
ov

c
,
w
h
e

c.Y

E .Y //

C
it is easy to deduce that Var. X

r
e

E .Y /// is minimized if we
take c

c.Y

N
X

Co
v.
N

N
:

V
ar
.

X;Y
/

Un
fort
una
tely
,
the
qua
ntit
ies
Co
v.
N

be

an
d
Va
r
N

N
Y

is

control
variate

fo

.Y /

ar
e
us
ua
ll
y
no
t
kn
o
w
n
be
fo
re
ha
nd
an
d
m
us
t

es
ti
m
at
ed
fr
o
m
th
e
si
m
ul
at
ed
da
ta.
T
he
qu
an
tit
y

ca
ll e
d
a

. and
To
se
e
w
hy
th
e
m
et

ar
e
po
sit
iv
el
y
co
rr
el
at
ed
.
If
a
si
m
ul
ati
on

ho
d
w
or
ks
,
su
pp
os
e
th
at

re (i.
su e.
lts N
in lar
a ge
la r
rg th
e an
va its
lu m
e ea
of n
its
m N.
ea A
n sim
E ilar
.
X
/
an
d
so
w
e
co
rr
ec
t
fo
r
th
is
by
lo
w
X

)
the
n
pro
bab
ly
als
o
N

E .Y /

is
la
rg
er
th
an

ar
gu
m
en
t
ho
ld
s
w
he
n

ar
e
ne
ga
ti
ve
ly
co
rr
el
at
ed
.

N and
In the simulation of the production line of section 2, a natural control variate would be the
long-term average production rate for the line with zero buffers.
Conditioning
The method of conditioning is based on the following two formulas. If X and Y are two arbitrary
random variables, then E . X / D E .E . X jY // and Var. X / D E .Var. X jY // C Var.E . X jY // Var.E
. X jY //. Hence, we conclude that the random variable E . X jY / has the same mean as and a
smaller variance than the random variable X .
How can we use these results to reduce the variance in a simulation? Let E . X / be the performance measure that we want to estimate. If Y is a random variable such that E . X jY D y/ is
known, then the above formulas tell us that we can better simulate Y and use E . X jY / than that
we directly simulate X .
The method is illustrated using the example of an M_=M_=1=N queueing model in which
customers who find upon arrival N other customers in the system are lost. The performance
measure that we want to simulate is E . X /, the expected number of lost customers at some fixed
time t . A direct simulation would consist of K simulation runs until time t . Denoting by Xi the
N D

P
K

number of lost customers in run i , then X . i D1 Xi /=K is an unbiased estimator of E . X /. However,

we can reduce the variance of the estimator in the following way. Let Yi be the total amount of time in
the interval .0; t / that there are N customers in the system in the i -th simulation
D
rate _, it follows that E . X
Y

run.
Since
customer
s arrive
accordin
g to a
Poisson
process
with

Y /=K

_Y

. Hence
an
improved
estimator
would be
N
_ ,
where
ND

i
.

N i D1

P
In Table 9 and Table 10 we compare, for _ D 0:5, _ D 1, D 3 and

D 1000,
the
estimator
s

and confidence intervals for E . X / when we do not, resp. do, use conditioning. We conclude that

33
# of runs
10
100
1000

mean st. dev.

33.10 1 0 . 0 6
34.09
9.21
33.60
8.88

95% conf. int.

[26.86, 39.33]
[32.28, 35.90]
[33.05, 34.15]

Table 9: Estimation of E . X / without using conditioning

# of runs
10
100
1000

mean
33.60
33.82
33.36

st. d e v.
7.16
6.91
6.57

95% conf. int.

[29.17, 38.05]
[32.46, 35.18]
[32.95, 33.76]

Table 10: Estimation of E . X / using conditioning

in this example the use of conditioning reduces the standard deviation of the estimator and hence
also the length of the confidence interval with a factor 1.3.