See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/251941590

Modeling and Simulation for Product Design Process

Article  in  SIMULATION: Transactions of The Society for Modeling and Simulation International · February 2013
DOI: 10.1177/0037549712451776

CITATION READS

1 667

1 author:

Dr. Hossein Arsham

Johns Hopkins University
151 PUBLICATIONS   983 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Sensitivity Analysis For Primal and Dual Degenerate LP Problems View project

All content following this page was uploaded by Dr. Hossein Arsham on 01 June 2014.

The user has requested enhancement of the downloaded file.

 Simulation

Simulation: Transactions of the Society for

Modeling and Simulation International
89(2) 178–191
Modeling and simulation for product Ó 2012 The Society for Modeling and
Simulation International
design process DOI: 10.1177/0037549712451776
sim.sagepub.com

Hossein Arsham

Abstract
Descriptive simulation finds the performance measure of a system, given a particular value for the input parameters.
Inverse simulation reverses this and attempts to find the controllable input parameters required to achieve a particular
performance measure. This paper proposes using a ‘stochastic approximation’ to estimate the necessary design para-
meters within a range of desired accuracy. The proposed solution algorithm is based on Newton’s methods using a
single-run simulation to minimize a loss function that measures the deviation from a target value. The properties of the
solution algorithm and the validity of the estimates are examined by applying them to a reliability system with a known
analytical solution.

Keywords
system design by simulation, design parameter setting, reliability system design, robust design, gradient estimation, pre-
scriptive modeling and simulation

1. Introduction In discrete event systems, Monte Carlo simulation is

usually needed to estimate J(v) for a given value v = v0.
Simulation continues to be the primary method by which By the law of large numbers
engineers obtain information about analysis of complex
stochastic systems, such as production assembly lines, X
n
flexible manufacturing systems, and reliability systems. J^ðv0 Þ = 1=n Z ðyi Þ, ð2Þ
Almost every evaluation of a stochastic system’s perfor- i=1
mance can be formulated as an estimation of an expected
value. Consider a system with continuous parameter converges to the true value, where yi, i = 1, 2, ., n, are
v ∈ V 4 R, where V is an open interval. Let independent, identically distributed random vector realiza-
tions of Y from f(y; v0), and n is the number of indepen-
J ðvÞ = EY jv ½Z ðY Þ, ð1Þ dent replications. The numerical result based on (2) is
only a point estimate for J(v) at v = v0. The numerical
be the expected performance measure, where Y is a random result based on (2) is a solution to a descriptive analysis:
vector with known probability density function (pdf), f(y; given the underlying pdf with a particular parameter value
v) depends on v, and Z is the performance measure. For v0, estimate the output function J(v0). The direct problem
example, in a reliability system, J(v) might be the mean is widely used in stochastic systems analysis. Now we
time to failure, Z is the lifetime of a system, Y is the life- pose the prescriptive analysis: given a target output value
time of the components, and v might be the components’ of the system and a parameterized pdf family, find an
mean lifetimes. In general, v is the shape or scale para- input value for the parameter that generates such an out-
meter of the underlying pdf. put. The solution to the design problem has potential
In systems analysis, we resort to simulation when Z is
either unknown or is too complicated to calculate analyti- University of Baltimore, Baltimore, MD, USA
cally. Simulation is needed to estimate J(v) for most sto-
Corresponding author:
chastic system. The principal strength of simulation is its Hossein Arsham, University of Baltimore, 1420 North Charles Street,
flexibility as a systems analysis tool for highly complex Baltimore, MD, 21201, USA.
systems. Email: [email protected]
Arsham 179

application in stochastic systems analysis and design. uses of single-run gradient estimation of the performance
Mathematical formulation of the design problem is as measure by Score Function method. This section includes
follows: an outline on the necessary tools for solving efficiently the
Given t, find v e V 4 R subject to J(v) = t, where parameters design setting problem. Section 4 formally for-
ð mulates the product parameters design target decision.
Based on this formulation, a recursive solution algorithm
J ðvÞ = EY jv ½Z ðY Þ = Z ð yÞf ðy; vÞdy, ð3Þ
for estimating the parameter design with desirable accu-
racy is presented. Section 5 illustrates the proposed method
where Z: Rm ! R is a system performance measure and Y on a construction of a reliability system with specific mean
e Rm is a random vector (or a truncated stochastic process) time to failure. Finally, Section 6 provides some conclud-
with pdf f (y; v). ing remarks and ideas for further research and extensions.
Extensive computer processing is needed to find a
design parameter value given a desired target for the per-
formance measure of a given system. The designer simu- 2. Decision-maker’s modeling
lates the process numerically and obtains an approximation environment
for that same output. The aim is to match the numerical
and experimental results as closely as possible by varying The modeling techniques are the way of representing the
the values of input parameters in the numerical simulation. systems. The model must be well tuned to the purpose it is
The most obvious difficulty in solving the design prob- intended. Since a model of a system is a representation of
lem is that one cannot simply calculate a straightforward the system that contains those elements that affect the
solution and be done. Since the output has to be matched objective of our decision, it is important to identify the
by varying the input, an iterative method of solution is most important elements and categorize them. An under-
implied. standing of the decision problem required in order to group
This paper proposes using a ‘stochastic approximation’ together entities of the decision model in the same cate-
to estimate the necessary design parameters within a range gory. The desired output usually determines the controlla-
of desired accuracy. The proposed solution algorithm is ble inputs. The input into a system can be classified either
based on Newton’s methods using a single-run simulation as controllable or uncontrollable, as Figure 1 illustrates.
to minimize a loss function that measures the deviation Time-horizons for the modeling review must be
from a target value. The properties of the solution algo- selected that are short enough so that the uncontrollable
rithm and the validity of the estimates are examined by inputs (or probabilistic knowledge of them) will not
applying them to a reliability system with a known analy- change significantly. Even for short time horizons one
tical solution. might consider the time discounting factor for future peri-
The remainder of this article is divided into five sec- ods. The output is often called performance measure (or
tions. In the next section we introduce the decision-mak- indicator) for the system. Figure 1 depicts the structured
er’s modeling environment. It follows by an outline on decision-making modeling process approach.

Figure 1. The modeling elements in the decision-maker’s environment.

180 Simulation: Transactions of the Society for Modeling and Simulation International 89(2)

It is a fact that in any organization, when an indicator is Action

used to measure performance then performance improves. Action is the ultimate decision for the product design
Moreover, when performance is measured and reported, parameter that achieved a desirable performance measure.
the rate of improvement accelerates. These parameters setting are determined after enough
Figure 1 depicts the structured decision-maker’s model- experiments are done with different set of controllable
ing that is the first step in the modeling and simulation inputs. This action is the strategy (i.e., purposeful deci-
process. Remember that, when structures and systems are sion) that results the desirable performance measure (e.g.,
aligned, they facilitate empowerment. When they are not, what is the number of cashiers during certain time period
they work against it. The following explain the element of of a day).
a model and the sequence by which they must be under- A detailed treatment of modeling construction processes
stood by the designer. and simulation approaches used in decision making from a
range of subject areas are given elsewhere.1–3
Performance measure (or indicator)
Measuring expected product performance is at the top of Scope and limitations
the designer’s concerns. The development of effective per- For clarification of the scope and limitation of this paper,
formance measure (or indicator) is seen as increasingly the following provides descriptions of the other similar
important by many in a competitive market. Performance phrases to avoid any possible confusion.
management is a cyclical process. It starts with effective
and efficient planning (e.g. the average waiting time of
customers in a local bank, compared with other neighbor- Reverse engineering
ing banks). Reverse engineering is the process of discovering the tech-
nological principles of a device, object, or system through
analysis of its structure, function, and operation. As an
Uncontrollable inputs example of reverse engineering, recently an unmanned US
These come from the decision-maker’s environment. drone fell into the hands of a hostile foreign government.
Uncontrollable inputs often create the problem and con- It is believed that it was sent to some aero-engineering
strain the actions (e.g. arrival rate of customers to a bank, lab, where their engineers attempted to figure out its parts,
at a certain time period of the day). its main functional structures, and their working
mechanisms.4

Parameters of the business

Feasibility problem and goal-seeking indicators
Parameters are the constant elements that do not change
during the time horizon of the decision review. These are In most business decision-making applications, the man-
the factors partially defining the decision problem (e.g. ager wishes to achieve a specific goal for the organiza-
total number of customers, different services that the bank tion’s overall objective, while satisfying the constraints of
provides, etc.). the model. The user does not particularly want to optimize.
This type of problem is usually called a feasibility prob-
lem. In such situations, the decision-maker aims at satisfy-
Controllable inputs ing or making incremental changes, rather than optimizing.
This is because the human mind has a bounded rationality
The collection of all possible courses of action the decision and hence cannot comprehend all alternatives. In the incre-
maker (i.e. the designer) might take (average service time mental approach to decision-making, the manager takes
determined by number of cashiers at different periods of only small steps, or incremental moves, away from the
time in a given day). existing system. This is usually accomplished by a ‘local
These inputs are the parameter design that we are con- search’ to find a ‘good enough’ solution. This problem is
cerned with. Their values determine the performance (i.e. referred to as a ‘satisfying problem’, ‘feasibility problem’,
life) of the product. or ‘goal-seeking’ problem. Therefore, the aim is to achieve
a global improvement to a level that is good enough, given
current information and resources. One reason that a
Interactions involving these components decision-maker manager overestimates the importance of
These are logical dependencies among all of the above the optimal strategy is that organizations often use indica-
components (e.g. the process that a customer goes through tors as ‘proxies’ for satisfying their immediate needs. Most
between entering and leaving the bank). managers pay attention to indicators, such as profit, cash
Arsham 181

flow, share price, etc., to indicate survival rather than as a general in nature, known as regularity conditions such as
goal for optimization. Another approach is to use ‘goal continuity, differentiability, invertability, and so on.
programming’ models that deal precisely with problems of The simulation models based on (2), although simpler
constraint satisfaction, without necessarily having a single than the real-world system, are still a very complex way
objective. Basically, they look at measures of constraint of relating input (v) to output J(v). Sometimes a simpler
violation and try to optimize them. analytic model may be used as an auxiliary to the simula-
tion model. This auxiliary model is often referred to as a
local response surface model (known also as a metamo-
3. Tools for setting production design del).7 Local response surface models may have different
parameters objectives: model simplification and interpretation,8 opti-
mization,9,10 what-if analysis,11 and generalization to mod-
The design problem is essentially backwards. The output
els of the same type. The following polynomial model can
is given, but the input must be determined. This is easiest
be used as an auxiliary model:
to appreciate when a designer wants to match experimen-
tal data in order to obtain some basic parameters. The
designer simulates the process numerically and obtains an J ðvÞ = J ðv0 Þ + dv:J 0 ðv0 Þ + ðdvÞ2 J 00 ðv0 Þ=2 + . . . , ð4Þ
approximation for that same output. The aim is to match
where δv = v – v0 and the primes denote derivatives. This
the numerical and experimental results as closely as possi-
local response surface model approximates J(v) for small
ble by varying the values of input parameters in the
δv. To estimate J(v) in the neighborhood of v0 by a linear
numerical simulation. Analyzing this, clearly, the output is
function, we need to estimate the nominal J(v) based on
there, and it is the input quantity that needs to be deter-
(2) and its first derivative. Traditionally, this derivative is
mined. The most obvious difficulty in solving the design
estimated by crude Monte Carlo, i.e. finite difference
problem is that one cannot simply calculate a straightfor-
which requires rerunning the simulation model. Methods
ward solution and be done. Since varying the input must
that yield enhanced efficiency and accuracy in estimating,
set the output, therefore an iterative solution method is
at little additional cost, are of great value.
implied. Our approach may be viewed as an optimization
There are few ways to obtain efficiently the derivatives
scheme where a loss function as described by Ross5 must
of the output with respect to an input parameter.12 The
be minimized. Therefore, the process of solving a design
most straightforward method is the score function (SF).
problem often comes down to finding the best method of
The SF approach13–15 is the major method for estimating
minimizing the loss function. The key part of optimization
the performance measure and its derivative, while obser-
is to compute the derivative of the output with respect to
ving only a single sample path from the underlying sys-
an input parameter.
tem.14 The basic idea of SF is that the derivative of the
In the case when v is any uncontrollable parameter, the
performance function, J0 (v), is expressed as expectation
designer is interested in estimating J(v) for a small change
with respect to the same distribution as the performance
in v = v0 to v = v0 + δv0. This is the so-called ‘what-if’
measure itself.
problem that is a direct problem. However, when v is a
This paper treats the design problem as a simulation (as
controllable input, the decision-maker is interested in the
opposed to regression) problem. By this approach, we are
product/service design parameters problem, i.e. ‘What per-
able to apply variance reduction techniques (VRTs) as
turbation of the design parameters (i.e. input parameters)
often used in the direct problem to enhance accuracy.
will achieve a desired change in the performance measure
Specifically, we embed a stochastic version of Newton’s
(i.e. the output value)?’ Another application of the design
method in a recursive algorithm to solve the stochastic
problem is where we may want to adapt a model to satisfy
equation J(v) = J for v, given J at a nominal value v0.
a new constraint with a stochastic function. While the
The explicit use of a linear local response surface model
what-if problem has been studied extensively, the design
is the target parameter design: given a desired value J =
parameter setting by simulation is relatively new. Design
J(v), find the prerequisite input parameter v.
interpolation based on regression models provides an indi-
Most engineering design methods essentially involve a
rect approach to solve the design problem. In this treat-
framework for arriving at a target value for product, pro-
ment, one simulates the system for many different values
cess, and service attributes through a set of experiments,
of v = v0, and then one approximates the response surface
that includes Monte Carlo experiments. To solve the prod-
function J(v).6 Finally, the fitted function is used to inter-
uct design problem, we will restrict our model to the first-
polate to obtain the unknown parameter v. Since the shape
order expansion. For a given J(v) the estimated δv using
of J(v) function is unknown, this approach is tedious, time-
(4) is
consuming and costly. Moreover, in random environments,
the fitted model might have unstable estimates for the  
^dv = J ðvÞ  J^ðv0 Þ =J^0 ðv0 Þ, ð5Þ
coefficients. The only information available about J(v) is
182 Simulation: Transactions of the Society for Modeling and Simulation International 89(2)

provided that the denominator in (5) does not vanish for H = S0 + S2: ð10Þ
any v0 in the interval V.
In the multidimensional case, the gradient and Hessian of
J(v) could be obtained in a straightforward manner by gen-
Gradient estimation by single-run methods eralizing these results.16 The estimator for the first and sec-
Simulation models, although simpler than real-world sys- ond derivatives based on (8) and (9) are given by
tems, are still a very complex way of relating input para-
0 X
n  
meters (v) to performance measures J(v). Sometimes a J^ ðv0 Þ = Z ðyi Þ:S yi ; v =n ð11Þ
simple analytical model may be used as an auxiliary to the i=1
simulation model. This auxiliary local response surface X
n
00
model is often referred to as a metamodel.7 In this treat- J^ ðv0 Þ = Z ðyi Þ:H ðyi ; v0 Þ=n ð12Þ
ment, we have to simulate the system for some different i=1
values of v and then use a ‘goodness-of-fit’ regression. We
where
fit a response surface to these data.6 Clearly, coupling the
simulation model with the SF method enhances the effi-
S ðyi ; v0 Þ = f 0 ðyi ; v0 Þ=f ðyi ; v0 Þ ð13Þ
ciency of local response surface model construction. A
local response surface model can also be constructed by and
using sensitivities in a Taylor expansion of J(v) in the
neighborhood of v = v0. The resulting local response sur- H ðyi ; v0 Þ = f 00 ðyi ; v0 Þ=f ðyi ; v0 Þ: ð14Þ
face model can be used for characterization (such as
increasing/decreasing and convexity/concavity) of the Note that both estimators (11) and (12) are evaluated at v
response surface. = v0, and yi are the same n independent replications used
For the readers’ convenience, we rewrite Equation (3) in (2) for estimating the nominal performance J(v0); there-
defined in the introduction here. Let fore, they are quite efficient in terms of the computation
cost. Estimates obtained by using (11) and (12) are
ð unbiased, consistent, and they converge to the true values
J ðvÞ = EY jv ½Z ðY Þ = Z ð yÞf ðy; vÞdy, ð6Þ in the sense of the mean squared error.16 The estimated
gradient can also be used in solving optimization problems
(where Z is a system performance measure and Y e Rm is a by simulation using the stochastic version of the classical
random vector (or a truncated stochastic process) with pdf nonlinear programming algorithms.17 Other applications
f(y; v)) be the performance measure, then of sensitivity information include stability analysis.12
ð The following section provides a descriptive presenta-
J 0 ðvÞ = ½Z ð yÞ:f ðy; vÞ0 dy, ð7Þ tion of other approaches to gradient estimations. For the
full algorithmic implementations and their interrelation-
where the prime (0 ) denotes the derivative with respect to ship, see Arsham12 and references therein.
v. Note that, despite the fact that y depends on v, only the
function Z.f is subject to differentiation with respect to v. Other approaches to gradient estimation
From (7) it follows that
In the design, analysis, and operation of discrete-event sys-
ð
tems, any information about the derivative dJ(v)/dv, is use-
J 0 ðvÞ = Z ð yÞf 0 ðy; vÞdy = EY jv ½Z ðY Þ:S , ð8Þ ful to both engineers and managers. The following
approaches avoid any numerical problems associated with
where S = f 0 (y;v) / f(y;v) is the SF. Differentiation is with the finite-differencing ratio as an approximation to the
respect to v. This is subject to the assumptions that the dif- derivative; they are based on a single simulation run, and
ferentiation and the integration operators are interchange- the methods have the potential for real-time applications.
able, f 0 (y;v) exists, and f(y;v) is positive for all v e V, where
V is an open interval. A necessary and sufficient condition
for the interchangeability used above is that there must be Finite difference approximation. Kiefer and Wolfowitz18
no discontinuity in the distribution.12 Similarly, the second proposed a finite difference approximation to the deriva-
derivative is tive. One version of the Kiefer–Wolfowitz (K-W) tech-
nique uses two-sided finite differences. The first fact to
ð notice about the K-W estimate is that it requires 2N simu-
J 00 ðvÞ = ½Z ðY ÞS 0 f ðy; vÞ + Z ðY ÞSf 0 ðy; vÞdy = EY jv ½Z ðY Þ:H  lation runs, where N is the dimension of the vector para-
ð9Þ
meter v. If the decision-maker is interested in gradient
estimation with respect to each of the components of v,
where then 2N simulations must be run for each component of v.
Arsham 183

This is inefficient. The second fact is that it may have a would like to realize this change by keeping the order of
very poor variance, and it may result in numerical calcula- events exactly the same. The perturbations will be so small
tion difficulties. that only the duration, not the order, of the states will be
affected. This effect should be observed in three succes-
Simultaneous perturbation methods. The simultaneous per- sive scenarios:
turbation stochastic approximation (SPSA) algorithm
introduced by Spall19–22 has attracted considerable atten- Step 1: How does a change in the value of a parameter
tion. There has recently been much interest in recursive vary the sample duration related to that parameter?
optimization algorithms that rely on measurements of only Step 2: How does the change in individual sample dura-
the objective function to be optimized, not requiring direct tion reflect itself as a change in a subsequent, particular
measurements of the gradient of the objective function. sample realization?
Such algorithms have the advantage of not requiring Step 3: Finally, what is the relationship between the
detailed modeling information describing the relationship variation of the sample realization and its expected
between the parameters to be optimized and the objective value?
function. For example, many systems involving complex
simulations or human beings are difficult to model, and
Harmonic analysis. Another strategy for estimating the gradi-
could potentially benefit from such an optimization
ent simulation is based on the frequency domain method,
approach. The SPSA algorithm operates in the same
which differs from the time domain experiments in that the
framework as the above K-W methods, but has the strong
input parameters are deterministically varied in sinusoidal
advantage of requiring a much lower number of simulation
patterns during the simulation run, as opposed to being kept
runs to obtain the same quality of result. The essential fea-
fixed as in the time domain runs. The range of possible val-
ture of SPSA which accounts for its power and relative
ues for each input factor should be identified. Then the values
ease of use in difficult multivariate optimization problems
of each input factor within its defined range should be chan-
is the underlying gradient approximation that requires only
ged during a simulation run. In time series analysis, t is the
two objective function measurements, regardless of the
dimension of the optimization problem (one variation of time index. In simulation, however, t is not necessarily the
basic SPSA uses only one objective function measurement simulation clock time. Rather, t is a variable of the model,
per iteration). The underlying theory for SPSA shows that which keeps track of certain statistics during each run.
the N-fold savings in simulation runs per iteration (per gra- Frequency domain simulation experiments identify the
dient approximation) translates directly into an N-fold sav- significant terms of the polynomial that approximate the
ings in the number of simulations to achieve a given relationship between the simulation output and the inputs.
quality of solution to the optimization problem. In other Clearly, the number of simulation runs required to identify
words, the K-W method and SPSA method take the same the important terms by this approach is much smaller than
number of iterations to converge to the answer, despite the the number of competing alternatives, and the difference
N-fold savings in objective function measurements (e.g. becomes even more conspicuous as the number of para-
simulation runs) per iteration in SPSA. meters increases.
SPSA can be seriously limited by, for example, the sta-
bility constraints of the system e.g. traffic intensity must
remain positive but less than one for a steady state of Some remarks on the various gradient estimation
dynamic system’s sensitivity estimation.10 approaches
Using the SF method, the gradient can be estimated simul-
Perturbation analysis. Perturbation analysis (PA)23–32
com- taneously, at any number of different parameter values, in
putes (roughly) what simulations would have been pro- a single-run simulation. The basic idea is that the gradient
duced, had v been changed by a ‘small’ amount without of the performance measure function, J0 (v), is expressed as
actually making this change. The intuitive idea behind PA an expectation with respect to the same distribution as the
is that a sample path constructed using v is frequently performance measure function itself. Therefore, the sensi-
structurally very similar to the sample path using the per- tivity information can be obtained with little computational
turbed v. There is a large amount of information that is the (not simulation) cost, while estimating the performance
same for both of them. It is wasteful to throw this informa- measure. It is well known that the crude form of the SF
tion away and to start the simulation from scratch with the estimator suffers from the problem of linear growth in its
perturbed v. In PA, moreover, we can let the change variance as the simulation run increases. However, in the
approach zero to get a derivative estimator without numer- discrete-event dynamic system, simulation the variance
ical problems. We are interested in the effect of a para- can be controlled by run length. Furthermore, information
meter change on the performance measure. However, we about the variance may be incorporated into the simulation
184 Simulation: Transactions of the Society for Modeling and Simulation International 89(2)

algorithm. A recent flurry of activity has attempted to the observed sample path, whereas SF requires no knowl-
improve the accuracy of the SF estimates. edge of the system other than the inputs and the outputs.
Now we look at the general (non-regenerative) case. In Therefore, when using SF methods, variance reduction is
this case, any simulation will give a biased estimator of necessary. The question is whether or not the variance can
the gradient, as simulations are necessarily finite. If n (the be reduced enough to make the SF estimator useful in all
length of the simulation) is large enough, this bias is negli- situations to which it can be applied. The answer is cer-
gible. However, as noted earlier, the variance of the SF tainly ‘yes’. Using the standard VRTs can help, but the
sensitivity estimator increases with increase in n; so, a most dramatic variance reduction occurs when using new
crude SF estimator is not even approximately consistent. VRTs such as conditioning, which is shown numerically
There are a number of ways to attack this problem. Most to have a mean squared error that is essentially the same
of the variations in an estimator come from the score func- as that of PA.
tion. The variation is especially high, when all past inputs
contribute to the performance and the scores from all are
included. When one uses batch means, the variation is 4. Parameters design setting problem
reduced by keeping the length of the batch small. formulation
A second way is to reduce the variance of the score to Most engineering system designs,33–35 such as product,
such an extent that we can use simulations long enough to process, and service design, involve a framework for arriv-
effectively eliminate the bias. This is the most promising ing at a target value for a set of experiments, which may
approach. The variance may be reduced further by using include Monte Carlo experiments. A random quality loss
the standard VRTs, such as importance sampling. Finally, function L(Zi) for a given system can be expanded in the
we can simply use a large number of independently identi- neighborhood of the target value t as follows:
cally distributed replications of the simulation.
PA and SF can be unified.12 Further comparison of the
LðZi Þ = L(t) + (Zi  t)L0 (t) + ðZi  tÞ2 L00 (τ)=2 + . . . :
PA and SF approaches reveals several interesting differ-
ences. Both approaches require an interchange of expecta- ð15Þ
tion and differentiation. However, the conditions for this It can be shown that L(Zi) converges in mean squared error
interchange in PA depend heavily on the nature of the prob- if |Zi – t| < 1 and derivatives are finite. Since the optimal
lem, and must be verified for each application, which is not loss is zero at t, Equation (15) reduces to the following
the case in SF. Therefore, in general, it is easier to satisfy SF
quadratic approximation
unbiased conditions. PA assumes that the order of events in
the perturbed path is the same as the order in the nominal
LðZi Þ = K ðZi  tÞ2 : ð16Þ
path, for a small enough change in v, allowing the computa-
tion of the sensitivity of the sample performance for a partic- In (16), K is some constant that can be determined in terms
ular simulation. For example, if the performance measure is of the customer’s tolerance limit (t – δv), which suggests
the mean number of customers in a busy period, the PA esti- that the product performs unsatisfactorily when Zi slips
mate of the gradient with respect to any parameter is zero. below this limit. Given that the cost to the customer is A
The number of customers per busy period will not change if dollars, then K = A/δv2. Without loss of generality, for sim-
the order of events does not change. plicity let K = 1.
In terms of ease of implementation, PA estimators may The aim of parameter design is to choose the setting of
require considerable analytical work on the part of the the design parameter v that minimizes the average loss (the
algorithm developer, with some ‘customization’ for each risk function). The risk function R(t) is the expected value
application, whereas SF has the advantage of remaining a of the loss function, which can be shown as
general, definable algorithm, whenever it can be applied.
Perhaps the most important criterion for comparison RðtÞ = EfLðZi Þg = ðJ  tÞ2 + VarðZi Þ: ð17Þ
lies in the question of accuracy of an estimator, typically
measured through its variance. If an estimator is strongly This risk function measures the average loss due to a prod-
consistent, its variance is gradually reduced over time and uct performance that is proportional to the square of the
ultimately approaches zero. The speed with which this deviation from the target value t, as shown in Figure 2. A
happens may be extremely important. Since in practice, parabolic representation estimates the quality loss, expressed
decisions normally have to be made in a limited time, an monetarily, that results when quality characteristics deviate
estimator whose variance decreases fast is highly desir- from the target values. The cost of this deviation increases
able. In general, when PA does provide unbiased estima- quadratically as the characteristic moves farther from the
tors, the variance of these estimators is small. PA fully target value. The acceptance range is between J(L) and
exploits the structure of discrete event systems and their J(U). Below the lower limit the product is rejected, and
state dynamics by extracting the needed information from above the upper limit, the product must be reworked.
Arsham 185

Risk Cost

Tolerance limit

Tolerance limit
Upper
Lower
K1 Reject cost K2 Rework cost

J(v)
J1 J(L) Y target J(U) J2

Figure 2. Tolerance concept in target design.

The parabolic curve shown in Figure 2 represents the where dj is any sequence of positive numbers satisfying the
Taguchi loss function. From the curve you can interpret following conditions:
that the amount of loss is minimal for the target (or nom-
inal value), and as you deviate from the target the amount ∞
X
of loss increases, even if you are within the specified lim- dj < ∞; ð21Þ
j=1
its of the process.
The non-adjustable variational noise, i.e. and

VarðZi jvÞ = VarðZi Þ, ð18Þ ∞

X
dj2 = ∞: ð22Þ
is a measure of variation among products. However, the j=1
role of product design is to reduce the (J − t)2 part of risk,
which is our interest in this paper. Note that all estimates The first condition is a necessary condition for the conver-
involved in computing δv based on (5), i.e. in gence δv to approach zero, while the second condition
asymptotically dampens the effect of the simulation ran-
  0 dom errors.36 These conditions are satisfied, for example,
^
dv = J ðvÞ  J^ðv0 Þ =J^ ðv0 Þ ð19Þ
by the harmonic sequence dj = 1/j. With this choice, the
are computed simultaneously from a single-run simulation rate of reduction of di is very high initially, but it may
of the nominal probability space (v = v0). Note that to esti- reduce to very small steps as we approach the root. Since
mate the derivative we do not need to rerun the simulation. simulation is partly statistical data generation, one per-
To estimate the derivative adds only moderate computa- forms simulation experimentation in order to generate
tional cost to the base simulation. ‘good’ data. Instead of classical dj = a/(a + j) with a = 1,
we have performed some pilot runs for integer values of
1 ≤ a ≤ 10, and found that for a = 9 one gets consider-
A recursive solution algorithm to parameter design able savings, in a number of iterations. Therefore, we have
problem used the better choice dj = 9 / (9 + j) for the application
presented in later. However, as always, one must be care-
The solution to the design problem is a solution of the sto-
ful in generalizing since we have used the choice dj for
chastic equation J(v) = J, which we assume lies in some
specific applications.
bounded open interval V. The problem is to solve this sto-
Although the feedback concept is not a simulation, but
chastic equation by a suitable experimental design to
a systemic concept, whatever paradigm we use one can
ensure convergence as δv approaches zero. The following
always incorporate feedback. For example, consider a dis-
algorithm involves placing experiment j + 1 according to
crete event system (DES) model that employs resources to
the outcome of experiment j immediately preceding it.
achieve certain tasks/processes, by only incorporating
That is,
decision rules regarding how to manage the stocks and
    thence how the resource will be deployed depending on
vj + 1 = vj + dj ½t  J^ vj =J^0 vj , ð20Þ the stock level; clearly, in the system structure there are
186 Simulation: Transactions of the Society for Modeling and Simulation International 89(2)

feedback loops. Usually when modelers choose an Step 4: RESET: Reset the seeds of random number genera-
approach to a DES, they often model the system as an tors to their initial values. Go to step 2.
open loop, or nearly open-loop system, making the system Note that, by resetting the seeds to their initial values,
behave as if there were no superior agent controlling the we are using the Common Random Variate approach14 as
whole production process. Closing the loops, as shown in a VRT.
Figure 3, should be an elemental task that the simulation
modeler should do. There must be awareness of system
behavior, particularly if the system is subject to human
The accuracy of the estimate
decision-making processes/activities. In the design problem, the input parameter is random,
The inverse simulation algorithm is based on an itera- while the output is fixed and given as a target value. Upon
tive method using differentiation and a feedback structure estimating the input parameter, we must provide a mea-
as shown in Figure 3. sure, such as a confidence interval, to reflect the precision
Since the adjustments are made in proportion to the of the estimate. To construct a confidence interval for δv
recent value, we must be sure that the results remain finite. using the estimator (19), let
This requires that J0 (v) does not vanish for v e V, where V is
an open interval. To prevent excessive over-correction, we Ai = J ðvÞ  Z ðyi ; v0 Þ, ð23Þ
assume further that the solution lies in some finite interval Bi = Z ðyi ; v0 ÞS ðyi ; v0 Þ, ð24Þ
V. Under these not unreasonable conditions, this algorithm
will converge in mean square; moreover, it is an almost sure and denote
convergence. For some generalizations and studies concern- X X
ing speed of convergence and acceleration techniques, see A= Ai =n and B = Bi =n, ð25Þ
Dippon and Renz.37 Finally, as in Newton’s root-finding
method,1,38 it is impossible to assert that the method con- then
verges for just any initial v = v0, even though J0 (v) may sat-
isfy the Lipschitz condition over V. Indeed, if the initial S 2 = S11 2  2vS12 + v2 S22 ð26Þ
value v0 is sufficiently close to the solution, which is usually
where
the case, then this algorithm requires only a few iterations to
obtain a solution with very high accuracy. X X
S11 = ðAi  AÞ2 =ðn  1Þ, S22 = ðBi  BÞ2 =ðn  1Þ,
ALGORITHM ð27Þ
Step 0: INPUTS
t = Desired output and
j = Iteration number X
vj = Controllable input parameter v S12 = ðAi  AÞðBi  BÞ=ðn  1Þ, ð28Þ
n = Sample size
U = Desired upper limit for absolute increment A 100(1 – a)% confidence interval for δv is given by
u = vj + 1 – vj
a = A desired significance level jdv  vj
P½n1=2 ≤ tn1, 1a=2  ≥ 1  a ð29Þ
S=B
Step 1: INITIALIZATION
Set j = 1
Set vj = v0
Step 2: ESTIMATIONS
J(vj) using (2)
J0 (vj) using (9)
Step 3: COMPUTATIONS
  0 
u = 9½t  J^ vj =½ð9 + jÞJ^ vj 

If |u| < U
Construct 100(1 – a)% confidence interval for
estimated v
Stop.
Otherwise
set vj + 1 = vj + u and j ! j + 1 Figure 3. Parameter design problem with a feedback loop.
Arsham 187

where tn-1,1-a/2 is the 100(1 – a/2) percentile of Student’s H ð yÞ = f 00 ðy; vÞ=f ðy; vÞ =
t distribution with n – 1 degrees of freedom.39  X  X 
2
v2 yi  8v yi + 12 =v2 , i = 1, 2, 3, 4:

5. An illustrative application: design of a The system lifetime is Z(Y1,Y2,Y3,Y4; v0) = max [min
reliability system (Y3,Y4), min (Y1,Y2)]. It is readily seen that the theoretical
expected lifetime of this system is J(v0) = 3/(4 v0).40 Now,
For most complex reliability systems, the performance
we apply our results to compute a necessary value for v to
measures, such as mean time to failure (MTTF), are not
obtain a particular value for J(v), say J(v) = 2. For this
available in analytical form. We resort to Monte Carlo
reliability system, the underlying probability density func-
simulation to estimate the MTTF function of the system.
tion is
The purpose of this section is a design problem: given a
specific desired target value for MTTF for a system find  X 
components’ life density function parameters to achieve f ðy; vÞ = v4 exp v yi , i = 1, 2, 3, 4: ð30Þ
the target value. A stochastic approximation algorithm is
used to estimate the necessary, controllable input para- The SF is
meters within a desired range of accuracy. The potential
X
effectiveness is demonstrated by simulating a reliability S ð yÞ = f 0 ðy; vÞ=f ðy; vÞ = 4=v  yi , i = 1, 2, 3, 4, ð31Þ
system with a known analytical solution.
Consider a coherent reliability subsystem consisting of H ð yÞ = f 00 ðy; vÞ=f ðy; vÞ =
four homogeneous elements, i.e. manufactured by an iden-  X  X 
2
tical process. Furthermore, the components have indepen- v 2
yi  8v yi + 12 =v2 , i = 1, 2, 3, 4: ð32Þ
dent random lifetimes Y1, Y2, Y3, and Y4, which are
distributed exponentially with rates v = v0 = 0.5.
The estimated average lifetime and its derivative for the
The first two elements, and the last two elements, are in
nominal system (v = v0 = 0.5) based on (2) and (9) are
series. These two series, each with two components, are in
parallel, as shown in Figure 4. X     
The system lifetime is Z(Y1,Y2,Y3,Y4; v0) = max [min J ðv 0 Þ = max min Y3, j , Y4, j , min Y1, j , Y2, j =n,
(Y3,Y4), min (Y1,Y2)]. It is readily seen that the theoretical ð33Þ
expected lifetime of this system is J(v0) = 3/(4 v0).40 Now
we apply our results to compute a necessary value for v to and
obtain a particular value for J(v), say J(v) = 2. For this X       
reliability system, the underlying pdf is J 0 ðv0 Þ = max min Y3, j , Y4, j , min Y1, j , Y2, j :S Yi, j =n,
 X  ð34Þ
f ðy; vÞ = v4 exp v yi , i = 1, 2, 3, 4: X       
J 00 ðv0 Þ = max min Y3, j , Y4, j , min Y1, j , Y2, j :H Yi, j =n,
The SF is ð35Þ
X
S ð yÞ = f 0 ðy; vÞ=f ðy; vÞ = 4=v  yi , i = 1, 2, 3, 4, Respectively, where Yi,j is the jth observation for the ith
component (i = 1, 2, 3, 4). We have performed a Monte
Carlo experiment for this system by generating n = 10,000
independent replications using SIMSCRIPT II.5 random
number streams 1–4 to generate exponential variates Y1,
Y2, Y3, Y4, respectively, on a PC. The estimated perfor-
mance is J(0.5) = 1.5024, with a standard error of 0.0348.
The first and second derivative estimates are –3.0933 and
12.1177 with standard errors of 0.1126 and 1.3321,
respectively.

A quadratic metamodel
The response surface approximation in the neighborhood
Figure 4. A reliability subsystem. v = 0.5 is
188 Simulation: Transactions of the Society for Modeling and Simulation International 89(2)

Table 1. A second-order polynomial local response surface model and direct simulation.

v Analytical Simulation Metamodel Relative error (%)

0.40 1.8750 1.8780 1.8723 0.14

0.42 1.7857 1.7885 1.7887 0.17
0.44 1.7045 1.7072 1.7098 0.31
0.46 1.6304 1.6330 1.6359 0.33
0.48 1.5625 1.5650 1.5667 0.27
0.50 1.5000 1.5024 1.5024 0.16
0.52 1.4423 1.4446 1.4430 0.05
0.54 1.3889 1.3911 1.3884 0.04
0.56 1.3393 1.3414 1.3386 0.05
0.58 1.2931 1.2951 1.2937 0.05
0.60 1.2500 1.2520 1.2537 0.29

Table 2. Iterative decision parameter estimate for the reliability system.

(1) Iteration number j (4) Estimated derivative

(2) Fixed input vj (5) Change in vj
(3) Estimated MTTF (6) New input parameter vj + 1
(1) (2) (3) (4) (5) (6)

1 0.5000 1.5024 − 2.9598 − 0.1513 0.349

2 0.3487 2.1544 − 6.0862 − 0.0208 0.369
3 0.3694 2.0333 − 5.4217 + 0.0046 0.374
4 0.3740 2.0083 − 5.2888 + 0.0011 0.375

J ðvÞ = 1:5024 + ðv  0:5Þð3:0933Þ P½0:374 ≤ v ≤ 0:377 ≥ 0:90: ð37Þ

+ ðv  0:5Þ2 ð12:1177Þ=2 + . . . Comparing the theoretical value v0 = 0.3750, obtained
≈ 0:6:0589v2  9:1522v + 4:5638 ð36Þ from J(v) = 3/4v0 = 2, with our computational value sug-
gests that the results based on the proposed algorithm are
A numerical comparison based on direct simulation and quite satisfactory. In fact, running this system with v =
the local response surface model (36) is given in Table 1. 0.375, and n = 10,000, we obtained an estimated MTTF of
The relative error as presented is the difference between J(v) = 2.0000. Hence, the discrepancy in the estimated
the metamodel and the analytical values. input parameter by this algorithm must be considered as a
Note that the largest error in Table 1 is 0.33% which pure random error which can be reduced by increasing n.
could be reduced by either more accurate estimates of the The metamodel (36) could also be applied to J(v) = 2 to
derivatives and/or using a higher-order Taylor expansion. estimate the desirable v. Solving the resulting quadratic
A comparison of the errors indicates that the errors are equation, the relevant root is v = 0.3725. This result is an
smaller and more stable in the direction of increasing v. inferior estimate for v compared with the iterative method,
This behavior is partly due to the fact that lifetimes are although the accuracy of the latter comes with greater
exponentially distributed with variance 1/v. Therefore, computational cost.
increasing v causes less variance than the nominal system
(with v = 0.50).
Now assume that the manufacturer wants to improve 7. Conclusions and some directions for
the average lifetime of the system to J(v) = t = 2. To
future research
achieve this objective, we have set v0 = 0.5 and U = 0.0001
in the proposed algorithm. The numerical results are tabu- Conventional approaches to simulation involve finding the
lated in Table 2. response of a system to a particular input or disturbance.
The estimated input parameter to achieve the output Inverse simulation reverses this and attempts to find the
J(v) = t = 2 is 0.375. A 90% confidence interval based on control input required to achieve a particular response. An
this estimate using (20) is inverse simulation method is introduced to be used in the
Arsham 189

analysis and design of systems. The methodology is pre- Some suggestions for future research topics
sented in the context of a reliability system application. We expect to introduce other efficient VRTs. The com-
This section includes a presentation of a solution algorithm mon random variates as a VRT are already embedded in
for the inverse simulation and issues of numerical stability the algorithm. Since
and accuracy. The methodology includes an iterative
method based on differentiation of the performance mea- ð ð ð
sure and use of feedback structures for generation of an E½S  = E½ln f 0 = ½ln f 0 f dx = f 0 dx = ½ f dx0 = 0,
inverse model, based on a stochastic version of Newton’s
ð38Þ
method.
Almost all DES simulation computation can be formu- we can express the gradient in terms of covariance between
lated as an estimation of an expected value of the system Z and S
performance measure, which is a function of an input para-
meter of the underlying pdf. In ordinary system simulation, J 0 ðvÞ = Cov½Z ðY Þ, S  = E½Z:S   E½Z :E½S  ð39Þ
this input parameter must be known in advance to estimate
and
the output of the system. From the designer’s point of
view, the input parameters can be classified as controllable
J 0 ðvÞ = E½Z ðY Þ:S  + αE½S  ð40Þ
and uncontrollable.2 The influential controllable input can
be recognized by factor screening methods.41 In this paper, where a could be the optimal linear control. Note also that
we considered the design problem: ‘What must be the per- (6) can be written as
turbation of the current controllable input parameter value ð
to achieve a desired output value?’ The approach used in
J 0 ðvÞ =Z ð yÞf 0 ðy; vÞdy
this study was as follows:
ð ð41Þ
= Z ð yÞ½f 0 ðy; vÞ=fðy; vÞfðy; vÞdy:
1. to estimate the derivative of the output function
with respect to the input parameter for the nominal
The best choice for f is the one proportional to Z(y).f 0 (y;
system by a single run and on-line simulation;
v). This minimizes the variance of J0 (v); however, this opti-
2. to use this estimated derivative in a Taylor’s
mal f depends on the performance function Z(y), which is
expansion of the output function in the neighbor-
not known in advance for most cases. One may use the
hood of the parameter; and, finally,
empirical version of Z(y).f 0 (y; v). We recommend a pilot
3. to use a recursive algorithm based on Taylor’s
run to study the effectiveness of these and other variance
expansion to estimate the necessary controllable
reduction techniques before implementing them.
input parameter value within a desired accuracy.
One may extend our methodology to higher-order
Taylor’s expansion. We believe that there is a tradeoff
Under some mild and reasonable conditions, the algorithm between number of iterations, sample size, and the order
converges to the desired solution with probability 1. The of Taylor’s expansion. Clearly, estimating the second deri-
efficiency of the proposed algorithm, in terms of accuracy, vative requires a larger sample size n, but fewer iterations
is tested using reliability in product designs with satisfac- to achieve the same accuracy.
tory results. The approach may have major implications We also expect to extend our methodology to the
for simulation modelers and practitioners in terms of time design problems with two or more unknown parameters
and cost savings. As always, since this experiment was by considering two or more relevant outputs to ensure
done on specific numerical example, one should be careful uniqueness. By this generalization, we could construct a
in making any other generalizations. linear system of stochastic equations to be solved simulta-
In this paper we have introduced the general concepts neously by multidimensional versions of the stochastic
of inverse simulation. An effective solution algorithm for approximation proposed elsewhere42,43 as well as the
the inverse simulation problem is presented from first Newton method17,44 using the second-order derivatives,
principles. The impact of the proposed inverse simulation e.g., Hessian.
method on conveying real understanding about the discrete The algorithms in this paper are presented in English-
event properties of the systems are now made available. like, step-by-step format to facilitate implementation in a
The inverse simulation method is also found to be of value variety of operating systems and computers, thus improv-
for the validation and control of complex DES models ing portability. However, there is a need to develop a deci-
with numerical stability and desirable accuracy. sion support system that makes the algorithms more
The proposed inverse simulation techniques can also be practically applicable to simulation in system design.42
applied as a measuring tool and decision procedure for the Another item for further research includes extension of
validation of simulation models. our experiment to large and more complex systems, such
190 Simulation: Transactions of the Society for Modeling and Simulation International 89(2)

as a large Jacksonian network with routing that includes 19. Spall J. Implementation of the simultaneous perturbation
feedback loops in order to study the efficiency of the pro- algorithm for stochastic optimization. IEEE Trans Aerospace
posed techniques. Electron Syst 1998; 34: 817–823.
20. Spall J. Adaptive stochastic approximation by the simulta-
Funding neous perturbation method. IEEE Trans Automatic Control
2000; 45: 1839–1853.
The National Science Foundation Grant CCR-9505732 supports
21. Spall J. Multivariate stochastic approximation using a simul-
this work.
taneous perturbation derivative approximation. IEEE Trans
Automatic Control 1992; 37: 332–341.
Acknowledgment 22. Spall J. A one-measurement form of simultaneous per-
I am most appreciative to the four reviewers for their careful turbation stochastic approximation. Automatica 1997; 33:
readings, useful comments and suggestions that are incorporated 109–112.
in the final version. 23. Cao X-R. Perturbation analysis of discrete event systems:
Concepts, algorithms, and applications. Eur J Operat Res
References 1996; 91: 1–13.
1. Chung Ch. Simulation Modeling Handbook: A Practical 24. Cassandras C. Discrete Event Systems: Modeling and
Approach. New York: CRC Press, 2003. Performance Analysis. Boston, MA: Irwin, 1993.
2. Morgan B. Applied Stochastic Modelling. New York: 25. Fu M. Optimization via simulation: A review. Ann Operat
Chapman and Hall/ CRC Press, 2008. Res 1994; 53: 199–247.
3. Samarskii A and Mikhailov A. Principles of Mathematical 26. Fu MC and Hu J-Q. Conditional Monte Carlo: Gradient
Modelling: Ideas, Methods, Examples. New York: CRC Estimation and Optimization Applications. Boston, MA:
Press, 2002. Kluwer Academic Publishers, 1997.
4. Eilam E and Chikofsky E. Reversing: Secrets of Reverse 27. Glasserman P. Derivative Estimation via Perturbation
Engineering. New York: John Wiley & Sons, 2008. Analysis. Boston, MA: Kluwer Academic Publishers, 1991.
5. Ross P. Taguchi Techniques for Quality Engineering. New 28. Glynn P. Likelihood ratio derivative estimation for stochastic
York: McGraw-Hill, 1996. systems. Commun ACM 1990; 33: 75–84.
6. Kleijnen J. Regression metamodel for generalizing simula- 29. Glynn P and Heidelberger P. Bias properties of budget con-
tion results. IEEE Trans Syst Man Cybernet 1979; 9: 93–96. strained simulations. Operat Res 1990; 38: 801–814.
7. Friedman L. The Simulation Metamodel. Boston, MA: 30. L’Ecuyer P. On the interchange of derivative and expectation
Kluwer Academic Publishers, 1996. for likelihood derivative estimation. Managem Sci 1995; 41:
8. Yu B and Popplewell K. Metamodels in manufacturing: A 738–748.
review. Int J Production Res 1994; 32: 787–796. 31. Vazquez-Abad F. Sensitivity analysis for stochastic DEDS:
9. Arsham H. Stochastic optimization of discrete event systems An overview. Aportaciones Matematicas Notas de
simulation. Microelectron Reliab 1996; 36: 1357–1368. Investigacion 1992; 7: 163–182.
10. Arsham H. Techniques for Monte Carlo optimizing. Monte 32. Yakowitz S, L’Ecuyer P and Vazquez-Abad F. Global sto-
Carlo Meth Appl 1998; 4: 181–230. chastic optimization with low-dispersion point sets. Operat
11. Arsham H. Gradient-based optimization techniques for dis- Res 2000; 48: 939–950.
crete event systems simulation. In Wah BW (ed.), Wiley 33. Clymer J. System design and evaluation using discrete event
Encyclopedia of Computer Science and Engineering. New simulation with AI. Eur J Operat Res 1995; 84: 213–225.
York: John Wiley & Sons, 2009, pp. 110–127. 34. Law A and Kelton W. Simulation Modeling and Analysis.
12. Arsham H. A simulation technique for estimation in per- New York, McGraw-Hill, 2000.
turbed stochastic activity networks Simulation 1992; 58: 35. Wild R and Pignatiello J Jr. An experimental design strategy
258–267. for designing robust systems using discrete-event simulation.
13. Kleijnen J and Rubinstein R. Optimization and sensitivity Simulation 1991; 57: 358–368.
analysis of computer simulation models by score function 36. Robbins R and Monro S. A stochastic approximation
method. Eur J Operat Res 1996; 88: 413–427. method. Ann Math Stat 1951; 22: 400–407.
14. Rubinstein R and Melamed B. Modern Simulation and 37. Dippon J and Renz J. Weighted means in stochastic app-
Modeling. New York: John Wiley & Sons, 1998. roximation of minima. SIAM J Control Optimiz 1997; 35:
15. Rubinstein R and Shapiro A. Discrete Event Systems: 1811–1827.
Sensitivity Analysis and Stochastic Optimization by the Score 38. Chen H and Schmeiser B. Stochastic root finding via retro-
Function Method. New York: John Wiley & Sons, 1998. spective approximation. IIE Transactions 2001; 33: 259–
16. Arsham H, Feuerverger A, McLeish D, Kreimer J and 275.
Rubinstein R. Sensitivity analysis and the ‘what-if’ problem in 39. Kleijnen J and Van Groenendaal W. Simulation: A Statistical
simulation analysis. Math Comput Modell 1989; 12: 193–219. Perspective. Chichester: John Wiley & Sons, 1992.
17. Polak E. Optimization: Algorithms and Consistent 40. Barlow R and Proschan F. Statistical Theory of Reliability
Approximations. New York: Springer, 1997. and Life Testing Probability Models. New York: Holt
18. Kiefer J and Wolfowitz J. Stochastic estimation of the Rinehart and Winston, 1975.
maximum of a regression function. Ann Math Stat 1952; 23: 41. Ruppert D. A Newton–Raphson version of the multivariate
462–466. Robbins–Monro procedure. Ann Stat 1985; 13: 236–245.
 Arsham 191

42. Benveniste A, Metivier M and Priouret P. Adaptive University of Baltimore. He received his higher-education
Algorithms and Stochastic Approximations. New York: degrees, all concentrated on modeling, from three different conti-
Springer-Verlag, 1990. nents. For more information, please visit his homepage at: http://
43. Wei C. Multivariate adaptive stochastic approximation. Ann home.ubalt.edu/ntsbarsh/index.html.
Stat 1987; 15: 1115–1130.

Author biography
Hossein Arsham is the Wright distinguished research professor
of decision science, statistics, and system simulation at the

View publication stats