This document discusses approaches to optimizing discrete-event simulation models over the past 20 years. While computational power has increased, recent literature shows a lack of new approaches and a widening divide between simulation modeling, optimization, and implementing improvements. The document proposes two areas for advancing the field: 1) integrating simulation optimization dynamically into operations rather than as a static tool, and 2) developing intelligent interfaces that can recognize input parameters and select appropriate optimization algorithms for specific problems.

1. Discrete-Event Simulation Optimization: A Review of Past Approaches
and Propositions for Future Direction
Linda Ann Riley, Ph.D.
Roger Williams University
School of Engineering, Computing and Construction Management
lriley@rwu.edu
Keywords: discrete-event simulation, optimiza-
tion, algorithmic optimization techniques
Abstract
Over the past twenty years, a significant body
of work has been undertaken on the topic of
methods and approaches to optimizing discrete-
event simulation models. Then, as is now, one of
the greatest challenges in optimizing discrete-
event simulations is the inability to precisely iden-
tify “the” optimal solution to a given system mod-
el. This is especially the case as the feasible so-
lution space expands.
Also over the past twenty years, computa-
tional speed has increased, computing and mod-
eling costs have decreased and theoretical de-
velopments in the field of simulation optimization
have emerged. Yet a divide appears to be widen-
ing. Recent literature indicates a lack of new, in-
novative approaches to optimizing large scale
discrete-event simulation models as well as an
absence in addressing the growing chasm be-
tween the simulation modeling, optimization and
outcome improvement processes. Many of the
studies and advances undertaken in the early to
mid-90’s are those still cited today when discuss-
ing simulation optimization.
This paper discusses and provides an over-
view of theoretical and methodological directions
in discrete-event simulation optimization. In addi-
tion, it suggests areas of study for advancing the
field. It is proposed that advances should move
the field of study and application in the direction
of blurring the boundaries between simulation
modeling, optimization and change implementa-
tion communities instead of widening the gaps.
1. INTRODUCTION
Academicians and practitioners have a num-
ber of tools to design, measure, study, analyze
and improve large, complex systems. One such
tool that has been utilized with especially good
results is simulation. Discrete-event simulation
specifically, allows for the realistic modeling of
stochastic events and the many process varia-
tions found in most complex, systems. One of the
primary outcomes of discrete-event modeling is
improvement in a system’s measures of perfor-
mance
Discrete-event simulation is a widely used tool
across many disciplines. Although each disci-
pline has system specific applications, the goal of
this technique usually involves system analysis
and/or performance improvement. By simulating
the dynamic nature of a system, one can better
understand and control random process varia-
tions. Furthermore, a good simulation that realis-
tically captures the system under study serves as
a model for experimentation. Since almost all
large-scale systems are both dynamic and sto-
chastic in nature, discrete-event simulation is an
excellent technique to study and analyze these
systems.
Algorithmic optimization approaches have
evolved over time as discrete-event simulation
has become more commonplace. Optimization
techniques involve numerous dynamic evalua-
tions of a simulation’s multi-dimensional solution-
space in the search for an optimal solution. Work
in the area of discrete-event simulation optimiza-
tion has concentrated for the most part on the de-
sign and evaluation of various algorithmic and
heuristic approaches in searching a simulation’s
solution space. At a higher level of abstraction,
there has been less of a focus on defining optimi-
780

2. zation frameworks. Although far outweighed by
the work in algorithmic development, work in the
area of optimization frameworks has been under-
taken by [Joshi et al. 1996, Abkay 1996], IEEE,
as well as the Department of Defense.
2. REVIEW OF DISCRETE-EVENT
SIMULATION OPTIMIZATION APPROACHES
Because computational time and cost are crit-
ical determinants of value and turn-around of the
simulation optimization process, a great deal of
ongoing work over the years has focused on ap-
plying the most appropriate algorithmic approach
considering the problem under study. Widely-
used search procedures for optimizing a simula-
tion’s feasible solution space include: determinis-
tic search methods; probabilistic search methods
and hybrid techniques.
Historically, a great deal of the literature in
discrete-event simulation optimization is based on
the probabilistic search techniques of: simulated
annealing [Liu 1999, Zolfaghari and Liang 1998,
Bailey et al. 1997, Haddock and Mittenthal 1992]
and evolutionary algorithms [Azadivar et al. 1999,
Hopper and Turton 1999, Pierreval and Tautou
1997]. A second area of algorithmic optimization
development has been in hybrid techniques. This
approach combines multiple algorithms into a
single optimization strategy [Shi et al. 1999,
Feyzbakhsh and Matsui 1999, Gong et al. 1997].
More recently, optimization strategies devel-
oped using evolutionary or nature-inspired algo-
rithms are referred to as metaheuristics [Glover
1986, Fu et al. 2005, Glover and Kochenberger
2003]. Metaheuristics provide a framework that
overcomes the need to customize an optimization
algorithm for different simulation problems. A
number of authors have discussed and explored
the theoretical underpinnings of metaheuristics as
well as various applications [Olafsson 2006, Yang
2010, Vasant 2012].
The primary reasons why metaheuristic algo-
rithms are particularly appropriate for discrete-
event simulation optimization are that these
methods: 1) handle both continuous and discrete
input parameters in contrast to search methods
requiring that input factors be expressed explicit-
ly; 2) deal well with conditions of local optima
compared to response surface methods; 3) re-
duce computational complexity in contrast to oth-
er search techniques, thus reducing solution iden-
tification speed, and; 4) perform quite well under
test conditions comparing a generated optimum
with complete enumeration of the solution space.
Other less applied algorithmic approaches for
optimizing discrete-event simulation models in-
clude particle swarm optimization [Clerc 2006,
Olsson 2011], honey bee algorithms [Nakrani and
Tovey 2006] and fire fly algorithms [Yang 2009].
Table 1, Overview of Commonly Used Dis-
crete-Event Simulation Optimization Approaches
and Algorithms presents the traditionally used
simulation optimization algorithms with brief
comments on the advantages, disadvantages and
processes involved in undertaking each. The ta-
ble also includes references to a sampling of the
seminal work in the area.
3. CHALLENGES WITH ADVANCING THE
KNOWLEDGE BASE OF DISCRETE-EVENT
SIMULATION OPTIMIZATION
One of the primary drawbacks of the system
modeling process is the lack of integration be-
tween the simulation model, the optimization pro-
cess and actions to enact system change as a
result of the optimization process. Ultimately the
goal of modeling many large-scale systems is to
increase the efficiency with which the system op-
erates as measured by the maximization or mini-
mization of selected parameters of the objective
function. Looking to the future, more emphasis
should be placed on blurring the boundaries be-
tween the simulation model, optimization and
change processes. To accomplish this goal, two
propositions are advanced in this paper.
1. Move from viewing discrete-event simulation
optimization as a static tool to one that is dynami-
cally integrated into operating practices.
For the most part, the present use of simula-
tion optimization in large-scale system simulation
scenarios is geared to problems that seek an op-
timal solution at time-specific points. Because of
the size of the models and the time required to
781

3. Table 1. Overview of Commonly Used Discrete-Event Simulation Optimization
Approaches and Algorithms
Method General Advantages General Disadvantages Process
Intuitive Methods  When used by an indi-
vidual familiar with the
system (expert), the
method can yield good
results.
 This method is a good
one to demonstrate the
concept of simulation
optimization in a teaching
environment.
 Computational time. 
Simulation time.
 No guarantee or confi-
dence that the ending solu-
tion is the optimal solution.
 Continuous variables are
problematic.
 Difficulty in selecting
both starting and stopping
points for the search.
 The user selects input parame-
ters and undertakes an iterative
process that involves: 1) varying
the parameter levels; 2) complet-
ing a statistically valid number of
simulation replications and runs,
and; 3) altering the input parame-
ters and reevaluating the results.
The objective of this method is to
find increasingly better solutions.
Complete Enumer-
ation
 Will produce the opti-
mal solution with small
models defined by a fi-
nite solution space.
 Computational time and
cost.
 Works only with discrete
variables.
 Wasted effort due to
testing every feasible solu-
tion in the feasible solution
space.
 Complete factorial experiment
of the model is undertaken. Anal-
ysis of all treatment combinations.
Tabu Search (see
[Lopez-Garcia et al.
1999, Glover 1977]
 Deals well with solu-
tion spaces character-
ized by local optima.
 Not well-developed as a
simulation optimization
methodology.
 Few studies comparing
accuracy and precision of
results.
 Works only for discrete
optimization models.
 Feasible solution space is ex-
plored by moving from one candi-
date to its best neighbor. Move-
ment occurs even if degradation in
the objective function is a result.
Tested solutions are considered
“tabu” for a user defined number
of iterations. Intensification and
diversifications strategies are
used to refine the search direc-
tion.
Pattern Search
(see [Findler 1987]
 Successful search
pattern transferrable to
similar simulation mod-
els.
 Does not deal well with
nonunimodality.
 Search moves in direction of
increasing improvement of the
objective function by “steps.” Step
sizes vary depending on the sen-
sitivity to change in the objective
function until a user-defined con-
vergence test or tolerance is satis-
fied.
782

4. Method General Advantages General Disadvantages Process
Genetic Algorithms
[ Hopper 1999, Col-
lins 1998, Aytug et
al.1998, Salzman
and Breitenecker
1995, Wellman and
Gemmill 1995,
Michalewiez 1994,
Goldberg 1994,
1989, Holland, 1975]
 Relatively fast com-
pared to other search
techniques.
 Interface process with
simulation models is
easy due to the design of
the algorithms.
 Does a good job at
identifying the global
optimum in models with
multiple local optima.
 Algorithms are exten-
sible.
 Robust method.
 Low computational
complexity.
 Good building block for
hybrid methods.
 Genetic algorithms can
be hard to analyze and
design depending on the
complexity of the manufac-
turing system being simu-
lated.
 Recognition of the need
for more theoretical work in
testing the accuracy of pro-
duced results.
 Based on the concept of evolu-
tion, genetic algorithms contain
three operators: selection, crosso-
ver and mutation. The search pro-
cess involves coding the parame-
ter set and searching a population
of points by means of probabilistic
transition rules. The search ends
when conditions of a termination
rule are met.
Simulated Anneal-
ing [ Liu, 1999,
Zolfaghari and Liang
1998, Bailey et
al.1997, Aarts and
Korst, 1989, Kirkpat-
rick et al. 1983]
 Technique is efficient
at moving from local op-
tima.
 Less computational
time required for each
search iteration however
more computational time
required overall because
more iterations usually
are needed.
 Low computational
complexity.
 Process avoids cy-
cling.
 Good building block for
hybrid methods.
 Process can require a
great deal of computation
time to find the optimal
solution.
 Attention must be paid to
the proper selection of a
seed solution or current
state starting point.
 Search process involves three
states: current state, neighboring
state and optimal states. At each
iteration, a change is made in the
current state and evaluated
against a neighboring state by
means of cost function. Transi-
tional probabilities and a tempera-
ture parameter dictate the likeli-
hood of moving from one state to
another. The search ends when a
user-defined number of iterations
or a user-defined number of opti-
mal states is achieved.
Hybrid Techniques
[ Alireza and Matsui
1999, Azadivar and
Tompkins 1999, Ma-
son et al. 1999, Fleu-
ry et al. 1999, Shi et
al. 1999, Chen and
Gen 1997, Ahmed et
al., 1998, Emelyanov
and Iassinovski
1997, Gong et
al.1997, Dolgui and
Ofitserov1997]
 Builds on established
successful algorithmic
procedures.
 Expected lower com-
putational complexity.
 Expected higher accu-
racy.
 Highly customizable
for specific scenarios.
 Usually designed to
handle both discrete and
continuous input parame-
ters quite well.
 Lack of algorithmic vali-
dation.
 Usually not extensible.
 Interface code can be-
come problematic depend-
ing on the hybrid technique.
 Customization can pre-
clude portability for other
manufacturing scenarios.
 Process is dependent on the
hybrid technique building blocks,
whether evolutionary strategies,
simulated annealing, deterministic
searches, or other.
783

5. run the model, the process is undertaken as a
static event in contrast to an integrated dynamic
process.
With the exception of some logistics optimiza-
tion applications involving transportation schedul-
ing, in most manufacturing and service settings,
optimization is not implemented as a dynamic tool
continuously running in the background and ulti-
mately driving certain operating decisions. To be
incorporated as a dynamic tool that contributes to
intelligent system design, work must continue in
further integrating the simulation modeling, opti-
mization and improvement implementation pro-
cesses. As optimization algorithms become more
sophisticated, the simulation optimization process
appears to be moving further away from the
modeling process. This is further exacerbated by
the three knowledge domains governing the three
processes. For the most part, the large-scale
system simulation modeling process is owned by
the industrial engineers, operation researchers
and simulationist community. Optimization algo-
rithms and frameworks are driven by the comput-
er science community and the improvement im-
plementation processes are owned by the effi-
ciency/managerial community. The optimization
black box is becoming more and more removed
from the ultimate modelers and especially users
of the simulation’s results. The gap appears to be
widening between research and theoretical de-
velopment in optimization approaches and appli-
cations in contrast to narrowing.
2. Need for simulation optimization procedures to
intelligently recognize input parameters.
When modeling large-scale system problems,
the use of off-the-shelf simulation packages is
many times necessary. Most of the most popular
discrete-event simulation packages have fast
learning curves, are graphically realistic, afforda-
ble, and produce easy to read, customized analy-
sis reports. In addition, add-on modules that allow
for external code-writing and customization are
common features of today’s off-the-shelf simula-
tion packages. In most cases also, these pack-
ages have built-in optimization modules.
One of the greatest drawbacks however of
these off-the-shelf simulation packages is the lack
of flexibility in altering the resident optimization
algorithms. Unless customized code is developed
external to the simulation package and then inte-
grated into the simulation, the user must take
what the vendor provides. Depending on the de-
sired optimization function and input parameters,
the vendor resident procedure may be wholly in-
adequate for the situation under study. Further-
more, discrete-event package vendors closely
guard as proprietary knowledge, the exact code
used to optimize their products. Selection of a
specific procedure should be dependent on
unique characteristics of the optimization problem
and not necessarily what is included in the off-
the-shelf simulation software. This is perhaps
one reason why the move toward metaheuristic
frameworks has occurred. These general pur-
pose approaches are evaluated as effective and
efficient over a range of problems.
To address this shortfall, some type of intelli-
gent interface is suggested. This interface could
be designed to choose from among a number of
algorithmic optimization procedures based on the
objective function and input parameters under
evaluation at any particular moment. This implies
perhaps an additional layer of AI/neural code that
could be incorporated into the optimization pro-
cess. Ultimately, this intelligent interface could
“learn” to recognize common optimization scenar-
ios, select starting and stopping rules, and poten-
tially also interface with the system improvement
framework.
As a further extension to the intelligent inter-
face, dynamic algorithmic visualization capabili-
ties could be incorporated into the optimization
procedures. Immersive technologies are used in
many simulation arenas. Incorporating immersive
visualization into optimization would serve to
bring a transparency between the modeling and
optimization processes. This would allow users
and decision makers to interactively view, and po-
tentially redirect the optimization process. In es-
sence, this feature would provide the decision
maker the ability to immerse him or herself into
the model, thus “directing” both the simulation
and optimization processes.
784

6. 4. CONCLUSION
As a tool to design and improve large-scale
systems, discrete-event simulation optimization is
ideally suited for addressing the complexity asso-
ciated with systems characterized as discrete-
event and stochastic in nature. With the variety
and robustness of algorithmic optimization proce-
dures, virtually all types of system problems can
be modeled and optimized. Yet, even though the
algorithmic development in optimization has been
well researched, there appears to be widening
gaps between the modeling, optimization and im-
plementation communities. Furthermore, due to
lack of transparency among the work of these
three groups, development of integrative frame-
works has been lacking. For optimization work to
advance in discrete-event modeling, it is pro-
posed that movement toward the design of dy-
namically integrated simulation/optimization/im-
plementation products be furthered explored. In
addition, intelligent optimization interfaces are al-
so proposed for off-the-shelf discrete-event simu-
lation packages. Advances in the area of dis-
crete-event simulation optimization should move
in the direction of blurring the boundaries be-
tween simulation modeling, optimization and
change implementation instead of widening the
gaps.
REFERENCES
Aarts, E. and Korst, J. (1989) Simulated Anneal-
ing and Boltzmann Machines: A Stochastic
Approach to Combinatorial Optimization and
Neural Computing. John Wiley and Sons,
Chichester, U.K.
Ahmed, M., Alkhamis, T. and Miller, D. (1998)
“Discrete Search Methods for Optimizing Sto-
chastic Systems.” Computers & Industrial
Engineering, Vol. 34, pp. 703-716.
Abkay, Kunter S. (1996) “Using simulation optimi-
zation to find the best solution.” IIE Solutions;
Norcross, Vol. 28, Issue 5 pp. 24-29.
Aytug, H., Bhattacharyya, S. and Koehler, G.
(1998) “Genetic learning through simulation:
An investigation in shop floor scheduling.”
Annals of Operations Research, Vol. 78 pp. 1-
28.
Azadivar, F. and Tompkins, G. (1999) “Simulation
optimization with qualitative variables and
structural model changes: A genetic algorithm
approach.” European Journal of Operational
Research. Amsterdam, Feb 16, 1999. Vol.
113, Issue 1. pp. 169-192.
Bailey, R.N., Graner, K.M. and Hobbs, M.F.
(1997) “Using simulated annealing and genet-
ic algorithms to solve staff-scheduling prob-
lems.” Asia-Pacific Journal of Operational Re-
search; Singapore, Vol. 14 Issue 2 pp. 27-43.
Chen, R. and Gen, M. (1997) “Parallel machine
scheduling problems using memetic algo-
rithms.” Computers & Industrial Engineering;
Vol. 33; pp. 761.
Clerc, M. (2006) Particle Swarm Optimization.
Wiley-ISTE.
Dolgui, A., Ofitserov, D. (1997) “A stochastic
method for discrete and continuous optimiza-
tion in manufacturing systems.” Journal of In-
telligent Manufacturing, Vol. 8 Issue 5; pp.
405-413.
Emelyanov, V., Iassinovski, S.I. (1997) “An AI-
based object-oriented tool for discrete manu-
facturing systems simulation.” Journal of In-
telligent Manufacturing, Vol. 8 Issue 1; pp.
49-58.
Feyzbakhsh, A. and Matsui, M. (1999) “Adam-
Eve-like genetic algorithm: a methodology
for optimal design of a simple flexible as-
sembly system.” Computers & Industrial
Engineering, Vol. 36 pp. 233-258.
Findler, N. V., Lo, C. and Lo, R. (1987) “Pattern
Search for Optimization.” Mathematics and
Computers in Simulation 29, no. pp. 41–50.
Fleury, G., Goujon, J., Gourgand, M., Lacomme,
P. (1999) “Multi-agent approach and sto-
785

7. chastic optimization: random events in
manufacturing systems.” Journal of Intelli-
gent Manufacturing, Vol. 10 Issue 1; pp. 81-
101.
Glover, F. and Laguna, M. (1997) Tabu Search,
Kluwer, Boston.
Glover F. (1986) “Future paths for integer pro-
gramming and links to artificial intelligence.”
Computers and Operations Research, Vol.13,
pp. 533-549.
Glover F. and Kochenberger G. A. (2003) Hand-
book of Metaheuristics, Springer.
Goldberg, D.E. (1989) Genetic algorithms in
search, optimization, and machine learn-
ing. New York, NY: Addison Wesley.
Goldberg, D. E. (1994) “Genetic and evolution-
ary algorithms come of age.” Association of
Computing Machinery Communications of
the ACM, Vol. 37, Issue 3, pp. 113-122.
Gong, D., Gen, M., Yamazaki, G. and Xu, W.
(1997) “Hybrid evolutionary method for ca-
pacitated location-allocation problem.”
Computers & Industrial Engineering, Vol.
33, pp. 577-80.
Haddock, J. and Mittenthal, J. (1992) “Simulation
optimization using simulated annealing.”
Computers & Industrial Engineering. Vol. 22
Issue 4; pp. 387-395.
Holland, J.H. (1975) Adaptation in natural and
artificial systems. MIT Press.
Hopper, E. and Turton, B. (1999) “A genetic al-
gorithm for a 2D industrial packing problem.”
Computers & Industrial Engineering, Vol. 37
Issue 1, 2; pp. 375-378.
Joshi, B.D., Unal, R. White, N.H., and Morris,
W.D. (1996) “A framework for the optimiza-
tion of discrete event simulation models.”
Presented at the 17
th
ASEM National Con-
ference, Dallas.
Liu, J. (1999) “The impact of neighborhood size
on the process of simulated annealing:
Computational experiments on the flowshop
scheduling problem.” Computers & Industrial
Engineering, New York; Vol. 27 Issue 1,2;
pp. 285-288.
Lopez-Garcia, L., and Posada-Bolivar, A. (1999)
“A simulator that uses Tabu Search to ap-
proach the optimal solution to stochastic in-
ventory models.” Computers & Industrial En-
gineering, Vol. 37 Issue 1,2; pp. 215-218.
Mason, A., Ryan, D. and Panton, D. (1999) “Inte-
grated simulation, heuristic and optimisation
approaches to staff scheduling.” Operations
Research, Vol. 46 Issue 2; pp. 161-178.
Michalewiez, Z. (1994) “Evolutionary compu-
tation techniques for non-linear program-
ming problems.” International Transac-
tions in Operational Research. Vol. 1 Is-
sue 2; pp. 233-240.
Nakrani, S and C. Tovey (2004) “On honey bees
and dynamic server allocation in Internet
hosting centers,” Adaptive Behavior, 12, 223-
240.
Olafsson, S. (2006) “Metaheuristics,” in Nelson
and Henderson (eds.) Handbook on Simula-
tion, Elsevier, 633-654.
Olsson, A., ed. Particle Swarm Optimization:
Theory, Techniques and Applications, Nova
Science Publishers Inc, 2011.
Pierreval, H. and Tautou, L., May. (1996) “Using
evolutionary algorithms and simulation for the
optimization of manufacturing systems.”
French Institute of Mechanical Engineering
(IFMA) Campus des Cezearx, B.P. 265.
F63175 Aubiere Cedex France.
Salzman, M. and Breitenecker,. F. (1995) “Ge-
netic algorithms in discrete event simulation,”
in Proceedings of the Eurosim. Congress 95.
September 11 - 15. Vienna. Austria. pp. 213-
218.
786

8. Shi, L., Olafsson, S. and Chen Q. (1999) “A new
hybrid optimization algorithm,” Computers &
Industrial Engineering, Vol. 36; pp. 409-426.
Shi, L., Olafsson, S. and Sun N. (1999) “New
parallel randomized algorithms for the travel-
ing salesman problem,” Computers & Opera-
tions Research, Vol. 26; pp. 371-9.
Vasant, P. (2012) Meta-Heuristics Optimization
Algorithms in Engineering, Business, Eco-
nomics, and Finance. Edited by Pandian
Vasant. 1st ed. IGI Global.
Yang, X. (2010) Engineering Optimization: An
Introduction with Metaheuristic Applications,
1st ed. Wiley.
Yang X. S. (2009) “Firefly algorithms for multi-
modal optimization,” 5th Symposium on Sto-
chastic Algorithms, Foundation and Applica-
tions (SAGA 2009) (Eds. Watanabe O. and
Zeugmann T.), LNCS, 5792, pp. 169–178.
Wellman, M.A. and Gemmill, D.D. (1995) “A
genetic algorithm approach to optimization
of asynchronous automatic assembly sys-
tems,” International Journal of Flexible
Manufacturing Systems, Vol. 7 Issue 1;
pp.27
Zolfaghari, S. and Liang, M. (1998) “Machine
Cell/Part Family Formation Considering Pro-
cessing Times and Machine Capacities: A
Simulated Annealing Approach,” Computers
& Industrial Engineering, Vol. 34, No. 4; pp.
813-823.
Biography
Dr. Linda Ann Riley is currently Engineering
Program Coordinator and Professor of Engineer-
ing for the School of Engineering, Computing and
Construction Management at Roger Williams
University (RWU). Previously, she held the posi-
tion of Associate Department Head for the De-
partment of Industrial Engineering at New Mexico
State University (NMSU). In addition, she served
as the founder and Director of the Advanced
Modeling and Simulation Laboratory at NMSU
and Director of a university-wide economic devel-
opment research center.
Dr. Riley has extensive business and engi-
neering consulting experience. As well, she is an
active researcher, teacher and author in the area
of simulation modeling and large-scale system
optimization. She has taught over 30 different
courses in her career, many of them simulation
focused and has written or co-authored over 120
academic/research publications and over 150 re-
search proposals.
Dr. Riley earned an M.S. in Industrial Engi-
neering as well as a Ph.D. in Logistics from New
Mexico State University, completed a two-year
post graduate fellowship at Brown University,
earned an MBA from Suffolk University and an
undergraduate degree from Boston University.
787