You are on page 1of 14

A Survey Of Optimization Techniques Being Used In The Field

Vivek D. Pinto, NCSA, University of Illinois at Urbana-Champaign


William M. Pottenger, Lehigh University, Bethlehem, PA and NCSA
William Tilt Thompkins, NCSA, University of Illinois at Urbana-Champaign

The rapid progress made in the application of optimization techniques to


industrial processes has been driven primarily by the increase in global
competitiveness and environmental regulations that force companies to
make optimum use of their resources. The US Air Force has of late been
coming under severe pressure to reduce expenditure in the form of
reductions in budgetary allocations. There has therefore, been an
increasing interest to optimize the operations of its various fleets. This
paper surveys some of the optimization techniques currently in use in an
attempt to analyze the strengths and weaknesses inherent in them and
their applicability to a military operation.

Optimization may be defined as the decision- making is hampered by the


process by which an optimum is presence of multiple objectives operating
achieved. The optimum may be that of on different time scales, strategic
an industrial institution or an objective environments and rates of change.
function that models a similar entity.
The factors that define the optimum will The US Air Force operates in a different
vary with the situation to which the competitive environment. However, it
optimization process is applied. Some has of late been coming under severe
examples of factors that may be pressure to reduce expenditure in the
optimized are cost, raw materials used, form of reductions in budgetary
time required and pollution caused. allocations. There has therefore, been an
Optimizations may be performed for increasing interest to optimize the
obtaining local and global optima. operations of its various fleets. To
determine the optimization technique
The increasing complexity of decision- that would be most effective for the Air
making that increased competition has Force business model, we surveyed
engendered has resulted in the some of the optimization techniques
implementation of enterprise-wide currently in use. We attempt to analyze
resource planning systems (ERP). These the strengths and weaknesses inherent in
information systems have provided a the different techniques and their
wealth of information that is staggering applicability to a military operation.
in terms of sheer size. To achieve a
competitive advantage firms must be This paper begins with an introduction to
able develop the means to use the some of the terms commonly used in the
information provided by these systems industry and military when dealing with
to improve their processes. However, optimization in section 1. Section 2
describes some of the techniques that are Markov processes: Simply defined a
used and describes a few recent Markov process is one in which the rate
applications of optimization in different of input and output are the same.
industries and research areas. Section 3 Processes are usually treated as Markov
evaluates the pros and cons of each processes because they resemble them or
technique. Section 4 summarizes the because they simplify the mathematics
paper and concludes. involved.

1. Introduction to Terminology Multi- Level Optimization: We define


Used multi- level optimization as being
Attrib ute: An attribute is any parameter optimization that takes place on two
that has significant impact on the different time scales or at different
behavior of a system that is being amounts of granularity. Usually the
optimized and is being controlled in an optimizer at the higher level makes use
attempt to achieve an optimum. of the outputs of the optimizer at the
lower level.
Constraint: A constraint is used to refer
to any condition that limits flexibility. In Multi-Objective Optimization: When the
this case the extent to which an optimization tries to achieve multiple
optimizer can vary the parameters that objectives, it is said to be multi-
influence the outcome is limited by the objective. Often in this type of
constraint. optimization, one or more of the
objectives are treated as constraints.
Objective Function: An objective
function (also called a cost function) 2. The Problem Being Tackled
spells out the relationship between the The planning and allocation of resources
outcome (objective) and the attributes in the Air Force has depended heavily on
mathematically. simulation software called LCOM
(Logistic Composite Model) that
Policy Level Objectives: These are provides estimates of outcomes based on
macro- level objectives imposed upon a a stochastic set of inputs. These inputs
system due to strategic or operational include labor, spares, capacity and
necessities. These are broad, intuitive policy requirements. An analyst will
and articulate the tradeoff between typically run multiple simulations with
different and sometimes antagonistic different scenarios and based on his/ her
objectives. past experience come up with an
educated guess about the optimal
Readiness: of a system is a measure of operating scenario. This scenario at
the number of units in the system that present takes on average anywhere from
can perform to specifications with a two to three weeks. The lead-time
given amount of certainty. This certainty needed for analysis decreases the
is usually measured as the probability of accuracy of the estimates. Further the
occurrence with a required confidence overhead imposed by the planning time
level (usually 95%). hinders the ability of the Air Force to
rapidly respond to changing
requirements of the environment.
Advances in the field of optimization The second derivative of the function is
and in supercomputer processing power then used to determine whether the
have made his problem more tractable in points so obtained are maxima or
terms of time. By relegating the core minima. Clearly, therefore, the objective
optimization work to a machine the function in classical & direct
analyst could be better utilized in value optimization must be continuous and
enhancing activities such as risk twice differentiable. This requirement is
analysis. Further by reducing the time generally not met in real world
needed to provide an estimate of the problems. Classical techniques however,
optimum it is hoped that the estimates have the advantage of being
themselves will be of greater utility in mathematically neat, simple to
decision- making. understand and easy to use. Due to the
dynamic and discontinuous nature of the
This project will be implemented in problem at hand classical techniques
three phases the proof of concept were clearly not of significant use. Some
phase, the validation phase and the classical techniques include Gaussian
implementation phase. The proof of elimination, Newton Raphson and Gauss
concept phase will involve the selection Seidal methods.
of two optimization methods that would
be most suited to handling the problem Simulated Annealing
at hand. It will involve building a new This technique draws its inspiration from
model of the entire C17 fleet operations, the annealing process that many
independent of LCOM, and the substances undergo while changing
optimization of the model with one of state. Typically, a substance, such as
the two methods chosen. The second iron for example, when heated gains
phase will involve the proof of concept energy. This energy is gradually
and validation of both the optimization dissipated due to cooling. The
methods. The third phase will include temperature is thus a measure of the
the application of the optimization disorder in the iron. As it cools its
techniques so developed to LCOM. molecules gradually, loose energy and
gain order. This process continues until
3. Techniques Used and Some the system can achieve thermodynamic
Recent Applications equilibrium. In the case of annealing a
The optimization problem can be stated piece of iron, thermodynamic
mathematically as equilibrium would occur when the
temperature of the iron is the same as the
x* ARGMAXx {f(x): g(x) = 0, xX}, temperature of the surroundings.
Simulated Annealing tries to mimic this
process and therefore gets its name.
where S = {x: g(x) = 0, xX} is the set
of possible solutions.
One of the difficulties in using
Simulated Annealing is that it becomes
Direct techniques are based on the
very difficult to choose the rates of
derivative method in which the gradient
cooling and the initial temperatures for
of an objective function is calculated and
the system that is being optimized. This
set to zero to obtain points at which its
occurs primarily because of the absence
value becomes maximum or minimum.
of any rules for selecting them. The provided accurate positions and relative
selection of these parameters depends on strengths for multiple sources.
heuristics and varies with the system that
is being optimized. Boudet et al, [Ref. 18] at the
Commissariat a lEnergie Atomique,
Industrial Application used simulated annealing to design thin
Simulated Annealing has been used for a film filters. The y assigned parameters
variety of optimization problems. These such as the number of layers, tolerances
range from portfolio optimization to for the target, incidence angles etc as
image processing. Belegundu and parameters for the simulated annealer.
Constans [Ref. 14] have used simulated Conditions like adhesion to the substrate
annealing to reduce noise pollution. were provided as constraints. The
They use a software called SOAR (Shell authors found that the simulated
Optimization for Acoustic Radiation) to annealer converged in every case and
predict the vibration and sound power. that the accuracy of the optimum
The shape of a structure is modified until obtained depended upon the parameters
it becomes a weak radiator. The chosen. They also report, interestingly
simulated annealing algorithm is used to that the solution needed no further
find the optimal distribution of mass refinement.
locations from a noise perspective.

Cederberg and Collins, [Ref. 15] at the Neural Networks


Naval Research Laboratory, have used There has been an explosion of interest
simulated annealing to optimize the in Neural Networks. They have been
solution of inverse geoacoustic applied to modeling as well as to
problems. The solutions are obtained interpretation. The major reason for the
with the help of self starter an efficient universal appeal and utility of neural
forward modeling tool. Geoacoustic networks is that they are universal
problems arise when a source and an approximators. Neural networks derive
array of receivers are separated by a their inspiration from biological neuron
water column with unknown parameters. systems and the brain. A neural network
These systems are used to probe the consists of a number of sub-units called
bottom of the ocean. neurons. These are interconnected in
parallel to form a network as shown
Greening at Datavision Computing below.
Services together with Zakarauskas and Input Hidden Layer
Dosso of the Defense Research Layer
Establishment Atlantic Canada, [Ref 16]
Output
used simulated annealing to localize the Layer
sources of acoustic signals. They used a
modified Bartlett processor, which
matched the measured acoustic fields
with computed values. The search for
the parameters was then carried out
using simulated annealing. They found
that this simultaneous localization
Hence, the application of neural nets to
A typical neuron consists of inputs, a optimization problems is easy to expect.
summing function, a limiting or
threshold function and outputs. These Fundamental to the utilization of neural
correspond to the dendrons, the nets to optimization problems is the
synapses, the cell bodies and the axons computation of the gradient of a cost/
of a biological neuron. A diagram of an objective function with respect to the
actual neuron and a artificial one is parameters being optimized. For a
shown below. constrained optimization, hard bounds
can be introduced either on the inputs or
on outputs depending on convenience.
Another technique of limiting these
problems is by using an exterior penalty
function. The penalty function adds an
additional function to the objective
function that increases the cost of those
regions that lie outside the constraints
thereby making them infeasible.

A multi-criterion optimization problem


is generally solved using the minimax
method, a weighting method, a Lp -norm
method, a goal programming method or
W1 more commonly an -constrained
W2 method. In the -constrained method one
.
. of the objectives is optimized while the
wn others are treated as constraints. In the
minimax method the optimization is
outputs
weights Sum (k) carried out on each of the objective
Inputs
functions and the minimum of the result
is taken as the optimum. In the
weighting method each objective is
assigned a scalar weight that signifies its
For neural networks to perform any relative importance to the other
useful function they must be trained. The objectives. The optimization problem is
most common method of training found then converted into optimizing the
has been the backpropagation method weighted sum of the different objective
[Ref 6 & 7] in which an error function is functions.
minimized by changing the weights of
each synapse in proportion to the error There have been several attempts to use
calculated. The error being the neural networks to optimize discrete
difference between some target or combinatorial problems. However, there
desired value and the actual value have been reports of success as well as
(output) obtained. The analogy between failure in obtaining optimal solutions
the error function in backpropagation using different methods developed. The
training and the objective function in
optimization is too obvious to be missed.
conflicting reports need to be resolved Genetic Algorithms
with further research. These algorithms draw their inspiration
from various hypotheses of biological
Industrial Applications evolution. Historically, such hypotheses
Lu and Markwood [Ref. 33] have have proposed that species evolve
developed and applied neural networks through a process of survival of the
in the industry to control the weight of fittest. A population of a species (set of
the coating applied in a hot dip coating possible solutions in this case) is created.
line of a steel mill. They report The members of this species are allowed
significant savings in terms of reduction to reproduce and recombine to produce
in coating weight, error between the new offspring. The fittest offspring are
target and actual coating weight and then selected to go on to the next stage
coating material used. They found that namely, recombining and producing new
neural networks outperformed regression offspring (or new solutions). The
models and resulted in a 5% saving of pseudo-code for this may be written as
Zinc. [from Ref 1]
Initialize the population
Veluswami at Mittel Corporation, Evaluate initial population
Canada, along with the Department of Repeat
Electronics, Carleton University used Perform competitive
neural networks to optimize high speed selection
interconnects of VLSI circuits with Apply genetic operators
electromagnetic simulation models.[Ref. to generate new solutions
32] They found that the models once Evaluate solutions in the
developed operate with minimal CPU population
time. They also demonstrate the savings Until some convergence criteria
in both CPU resources and advantages is satisfied
that neural networks have over existing There are a number of optimization
techniques. techniques in the same family as genetic
algorithms. They include genetic
Yalcinoz and Short, [Ref. 31] report programming and evolutionary
having used neural networks for solving strategies.
dispatch problems in transmission
systems. They found that the results that Industrial Applications
they obtained were very close to those Duponcheele and Tilley used genetic
obtained using the quadratic algorithms to optimize the shape of an
programming method. automotive structural bumper beam.
[Ref. 27] The used a variant of genetic
Ko and Cho [Ref 31], report using neural algorithms called messy genetic
networks to optimize face milling algorithm. [Ref. 28] They carried out
operations. They used two neural their evaluations in collaboration with
networks one for estimating tool wear ECIA, a car component manufacturer.
length and the other for mapping input The aim of a designer is to increase the
and output relations from the data during moment of inertia of the cross-section
cutting. with respect to bending and thus increase
the absorption of kinetic energy. The
optimization involves increasing this environment. The correct behavior or
cross-section by increasing the mass of solution is discovered over time by
the beam and therefore its cost or performing a given action, determining
changing the shape of the cross-section. its consequences and using those
The authors found that the performance consequences in performing future
of the messy genetic algorithm's design actions. Hence, tabu search must keep a
was consistently better than the existing track or list of the paths it has
ones. traversed. Tabu search has two main
stages. In the first stage it explores the
Lee and Horner [Ref. 29] use group state space coarsely and determines
synthesis to model piano tones. A probable solutions. In the second stage,
genetic algorithm is used by them to it searches in the environs of these
select nearly contiguous groups. probable solutions to determine which
Independent frequency deviations are solution is the best. A detailed
provided to each group to simulate investigation on this algorithm can be
partial stretching. They found that the found in [Ref. 21]. Vujic et al compared
data required was reduced by half when the performance of tabu search to the
compared to additive synthesis. MonteCarlo method and found that tabu
Furthermore, the perceptual identity of search was superior with respect to both
sound was also preserved. the computational effort and the value of
the objective function.
Higuchi and Kajihara report [Ref. 30]
that MITI, Japan is developing Industrial Applications
Evolvable Hardware chips (EHW) as Lee and Ellis [Ref. 17] at the Graduate
part of its real- world computing project. Institute of National Research and
EHW is based on the idea of combining management and John Hopkins
hardware with genetic algorithms to University respectively, provide some
"execute reconfiguration" automatically. empirical comparisons of algorithms.
These chips are used in They applied tabu search and other
telecommunication equipment where algorithms to optimize monitoring of
they process compressed signals. They network design to the Taiwan power
report that the GA is invoked each time system. They report that the primary
the prediction performance of an EHW objective in monitoring network design
decreases due to changes in the data to is to maximize information acquisition
be compressed. The GA searches for a while minimizing cost. Accordingly they
better structure and autonomously measured the performance of the
reconfigures the hardware to changing algorithms based on the computational
environments. Based on their experience time or the cost and the effectiveness or
they conclude that the EHW concept frequency of obtaining good solutions.
with GAs can be applied to a wide The algorithms they considered were the
variety of applications. polytope method, simulated annealing
and genetic algorithms. They found that
Tabu Search simulated annealing and tabu search
This technique is based on the idea that were superior in performance to the
humans behave in a seemingly random other methods.
manner given the same learning
Chang and Wen report [Ref. 24] that requires that the objective function be
they had developed a mathematical estimated. Furthermore, the models that
model for trouble call analysis and are used to estimate the objective
used tabu search to solve this problem. function may be stochastic due to the
They tested their solution on a sample dynamic and random nature of the
distribution system and found that the system and processes.
tabu search based method was efficient.
Stochastic approximation techniques are
Tabu search has also been applied to not new. They had been first developed
scheduling m identical parallel by Robbins and Munro in 1956.[Ref 2,
processors with 'sequence dependent 8] Their development however, has been
setup times' to minimize the total slow due to their dependence on
execution time [Ref. 25]. Their simulation and the accompanying
scheduling had three phases. In the first computing power required.
phase unassigned jobs were assigned to
the processor. In phase 2 tabu search was The basic idea behind the stochastic
used to improve upon the starting approximation method is the gradient
solution produced in phase 1. In phase 3 descent method. Here the variable that
final adjustments were made to improve the objective function is to be optimized
the solution obtained. The authors upon is varied in small increments and
reported that this methodology resulted the impact of this variation (measured by
in producing "good quality solutions in the gradient) is used to determine the
reasonable running times" when direction of the next step. The magnitude
compared with an exact procedure. of the step is controlled to have larger
steps when the perturbations in the
He and Kusiak report [Ref. 26] that they system are small and vice versa.
used tabu search to configure an
assembly system. They indicate the x1 , x 2 ,.... x n I ...........(i )
production of modular products requires
a single assembly line. The assembly Stochastic approximation algorithms
system must be designed to improve c i x k + ci + 1 x k + 1 > L ... .... .. (ii . .. i n )
agility and cost efficiency. They set the based on various techniques have been
aspiration value to the current minimum developed recently. They have been
and use a stopping criterion of a fixed applied to both continuous and discrete
number of iterations. They found that objective functions. [Ref 3 & 4]
their heuristic algorithm that used tabu Recently, their convergence has been
search outperformed the general line proved for the degenerate case as well
balancing method. [Ref 9]. In their paper, LEcuyer et al
[Ref 5] illustrate the behavior of
Stochastic Approximation stochastic approximation with different
While some non-classical optimization estimation techniques.
techniques are able to optimize on
discontinuous objective functions, they Industrial Applications
are unable to do so when the complexity Simha R., and Kurose F., [Ref 19] with
of the data becomes very large. In this funding from DARPA and the Office of
case the complexity of the system Naval Research used stochastic
approximation to minimize the call setup had about 25,000 pilots and flight
time. They attempted to achieve this by attendants. Their yearly cost was
balancing the load on call processors. estimated at around $1.3 billion. The
They compare the performance of two authors reported that the use of MIP in
stochastic approximation techniques their optimizatio n code (named TRIP)
when applied to a circuit switched or a and subsequent enhancements resulted in
packet switched network. savings of $20 million per year over the
period 1985-90 and $3-5million from
1990-92. The code implemented in
Mixed Integer Programming(MIP) Fortran and C used an existing IBM
Integer programming is used for optimization subroutine library.
optimizing linear functions that are
constrained by linear bounds. Quite In his paper [Ref. 11] Ciriani found that
often, the variables that are being varied MIP when used to optimize the
can have only integer value (e.g. in modeling techniques used at IBM
inventory problems where fractional 'significantly improved performance'. He
values such as the number of cars in reported using MIP as a preprocessor
stock are meaningless). Hence, it is more and found that combinatorial problems
appropriate to use integer programming. with a flat objective function seemed to
Mixed integer programming is a type of benefit from heuristics. He also
integer programming in which not all of concluded that model formulation can
the variables to be optimized have provide a better integer polytope
integer values. Due to the linear nature approximation if methods like
of the objective function it can be coefficient tightening are used.
expressed mathematically as Dillenberger and Wollensak [Ref 13]
report that the use of mathematical
n
min C j X k
programming techniques (such as MIP)
j , k =1
in the decision support system of IBM's
Sindelfingen plant was successful.

Ballintjin, [Ref. 12] at Shell Laboratories


reports that MIP models that were used
where C is the coefficient matrix and X
to control mode switching at acceptable
is the attribute vector of attributes levels generated attractive schedules.
x1 .xn . Also (i) and (ii..in ) are the
However, the solution times obtained
constraints in the equation above. were not satisfactory.
Typically, MIP problems are solved by
using branch and bound techniques to
increase speed. 4. The Pros and Cons of Each
Technique
Industrial Applications While classical techniques would
Mixed Integer programming has been obviously not be applicable to the
used extensively at American Airlines. problem we have at hand, stochastic
In their paper [Ref. 10], Ranga et al, at techniques seem, intuitively, to be
American Decision Technologies, appropriate for our purpose. The
describe the optimization of crew literature survey demonstrated that each
pairings. American Airlines, in 1992, technique had its strengths and
weaknesses. It also demonstrated that the random mutation may be superior to
performance of each algorithm would be GAs in some cases [Ref. 34]. The lack of
heavily dependent on the nature of the proofs demonstrating the ergodicity of
problem itself and the heuristics that we GAs is one factor that makes this
used. technique unsuitable for our problem.

The major strengths of simulated Tabu search had the advantage of not
annealing, [Ref. 35] are that it can using hill-climbing strategies. Its
optimize functions with arbitrary degrees performance could also be enhanced by
on non-linearity, stochasticity, boundary branch and bound techniques. However,
conditions and constraints. It is also the mathematics behind this technique
statistically guaranteed of finding an was not as strong as those behind neural
optimal solution. However, it has its networks or simulated annealing.
disadvantages too. Like GAs it is very Furthermore, a solution space would
slow, its efficiency is dependent on the have to be generated. Hence, tabu search
nature of the surface it is trying to would require a knowledge of the entire
optimize and must be adapted to specific operation at a more detailed level. Battiti
problems. The availability of and Tecchiolli [Ref. 22] had compared
supercomputing resources, however, Simulated Annealing and Tabu Search
mitigates these drawbacks and makes on the Quadratic assignment problem.
simulated annealing a good candidate. They found that tabu search does require
extra overhead in terms of memory
Neural networks have the advantage that usage and adaptation mechanisms
the entire operations process can be compared to Simulated Annealing.
treated as a black box. This would ease However, it avoids the "traps" inherent
the burden of having to model the entire in simulated annealing such as attraction
system. It has, however, the basins and hence did work better in their
disadvantage of requiring us to gather case. Paulli J., [Ref 23] reports, on the
data and training the network. other hand, that simulated annealing is
Furthermore, the performance of the significantly better than tabu search. He
optimizer would be heavily dependent cites the excessive memory overhead as
on the quality of the data used. well as the time required by tabu search
to form a search trajectory as being the
Genetic algorithms perhaps seem to be main causes of the poor performance.
the most popular algorithms at present.
Their advantage lies in the ease of Stochastic Approximation did not have
coding them and their inherent as many applications reported as the
parallelism. The use of genotypes other techniques. This could have been
instead of phenotypes to travel in the because of various factors such as the
search space makes them less likely to lack of a metaphorical concept to
get stuck in local minima. They have, facilitate understanding and proofs that
however, certain drawbacks to them. are complex. It has recently shown great
GAs require very intensive computation promise, however, especially in
and hence are slow. They are also not optimizing non-discrete problems. The
guaranteed to give an optimal solution. stochastic nature of our model along
There are examples to show that simple with the complexity of the application
domain makes this an attractive version is most desirable. The
candidate. implementation must in no way
compromise the security of the
Mixed integer programming was found underlying hardware and software
to have the widest application. It was systems.
preferred to routing airline crews and
other similar problems that bore a close While the authors appreciated the
resemblance to the problem we had at strengths of each technique, Stochastic
hand. Furthermore, the mathematical Approximation and Simulated
rigor we were looking for was well Annealing best fit our application. We
established. However, as the nature of chose simulated annealing due in part to
our problem is continuous and dynamic the fact that it lent itself naturally to
we preferred to use either Simulated optimizing a scenario-based simulation
Annealing or Stochastic Approximation. [Ref 20]. Another reason that influenced
our choice included the fact that the
5. Summary and Conclusions mathematics behind Simulated
Annealing is based on the rigorous
This paper has surveyed a wide range of theory of Markov chains. Additionally,
applications to determine which several convergence proofs have existed
optimization technique has proven to be for it in technical literature for quite a
most successful. This knowledge will be few years. Finally, Simulated Annealing
applied in choosing one of these variants such as Adaptive Simulated
techniques to optimize a simulation Annealing and Quenching provide a
model of the operations of the Air Force. quick first pass solution. More
The paper has briefly enumerated the elaborate simulated annealing techniques
strengths and weaknesses of each such as enhanced simulated annealing
technique and elaborated in some detail [Ref. 36] provide a clean mechanism
the techniques themselves. with parallelism inherent in it to achieve
speed up.
Due to the very nature of the application,
the optimization methodology would
have to meet certain requirements. It 6. References
must be able to provide a ninety- five 1] Perry Gray, William Hart, Laura
percent statistical confidence level for its Painton, Cindy Phillips, Mike Trahan,
estimates of the global optimum. John Wagner, Sandia National
Second, it must be fault tolerant and able Laboratories Albuquerque, NM 87185
to produce realistic results even in the A survey of Global Optimization
event of minor inconsistencies in the Methods.
input set. Third it should be able to come http://www.cs.sandia.gov/opt/survey/
up with a good approximation of, at a
minimum, the local minima when 2] Robbins H., S. Munro, A stochastic
constrained to execute in a given time approximation method, Ann. Math. Stat.
frame. Fourth, it should be robust 22, 400-407.
enough to deal with noisy inputs and
underlying stochastic models. Lastly, the 3] Andradottir S., A new algorithm for
ease of implementing a highly secure stochastic approximation; Proc. Of the
Winter Simulation Conf., IEEE press
(1990) 364-366 12] Ballintijn K., "Optimization in
Refinery Scheduling: Modeling and
4] Andradottir S., A method for Solution"; Optimization in Industry,
Discrete Stochastic Optimization; Mathematical Programming and
Mgmt Sc., Vol 41, No12, Dec 95 Modelling Techniques in Practice, John
Wiley and Sons, 1993, pp191-199.
5] LEcuyer P., Giroux N., Glynn P;
Stochastic Optimization by Simulation: 13] Dillenberger C., and Wollensak A.;
Numerical Experiments with the M/M/1 "Demand Allocation for Capacity
queue in Steady state; Mgmt Sc.; Vol. Optimzation: A Solution in IBM
40, No. 10, Oct. 94. manufacturing"; Optimization in
Industry, Mathematical Programming
6] Pinto V., Application of Temporal and Modelling Techniques in Practice,
Neural nets to EKG interpretation; John Wiley and Sons, 1993, Vol. 2.pp.
Maters Thesis, Univ. of Tulsa, OK 96 90-98.

7] Haykin S., Neural Networks A 14] Belegundu, A., Constans E.;


Comprehensive Foundation; Prentice Minimizing Radiated Sound Power
Hall International Inc., ISBN: 0-13- from Vibrating Shells Theory and
908385-5. Experiment; Conference on
Optimization in Industry - 1997;ASME
8] Wasan, M.T., Stochastic pp.7-10.
Approximation, Cambridge University
Press, 1969 15] Cederberg R. J., Collins M. D.;
Application of an Improved Self Starter
9] Chen, H., Convergence rate of to Geoacoustic Inversion; IEEE Jour.
Stochastic Approximation Algorithms in Of Oceanic Engineering; Vol. 22, No. 1,
the Degenerate Case, SIAM Jour. Jan. 1997 pp. 102 -109.
Control Optim., Vol. 36, No1, pp. 100-
114, Jan 98 16] Greening M., Zakarauskas P., and
Dosso S.; Matched Field Localization
10] Anbil R., Barnhart C., hatay L., for Multiple Sources in an Uncertain
Johnson E., and Ramakrishnan V.; Environment, with Application to Artic
"Crew-pairing Optimization at American Ambient Noise; Jour. Of the Acoustic
Airlines Decision Technologies"; Society of America Vol. 101, No. 6,
Optimization in Industry, Mathematical June 1997 pp. 3525-3537.
Programming and Modelling Techniques
in Practice, John Wiley and Sons, 1993, 17] Lee Y., Ellis J. H.; Comparison f
pp 31-36. Algorithms for Nonlinear Integer
Optimization: Application to Monitoring
11] Ciriani Tito, "Model Structures and Network Design; Journal of
Optimization Strategies"; Optimization Environmental Engineering, June 1996
in Industry, Mathematical Programming pp. 524 - 529.
and Modelling Techniques in Practice,
John Wiley and Sons, 1993, pp. 263-271
18] Boudet T., Chaton P., Herault L., Scheduling Problem with Sequence
Gonon G., Jouanet L., Keller P.; Thin Dependent Setup Times"; Intl. Jour. of
Film Designs by Simulated Annealing; Production Econ., Vol. 43, Jun 1, 1996,
Applied Optics, Vol. 35, No. 31, Nov. pp. 79-99.
1996, pp. 6219-6219-6222.
26] He D. W., Kusiak A.; " Design of
19] Simha R., Kurose J. F.; On- line Assembly Systems for Modular
Minimization of Call Setup Time via Products"; IEEE Transactions on
Load Balancing: A Stochastic Robotics and Automation, Vol. 13, No.
Approximation Approach" IEEE 5, Oct 1997. Pp. 646-655.
Transactions on Communications, Vol.
42, No. 2/3/4, Feb/Mar/Apr. 1994 pp. 27] Duponcheele G., Tilley D. G.;
1141-1144. "Topological Optimization of a Bumper
Beam via The Messy Genetic
20] Pinto V., Pottenger W., Thompkins Algorithm"; Proc. of The Instn. Of
T., A MultiLevel, MultiObjective Mech. Engrs., Vol. 212, Part D, pp. 133-
Optimization of a Stochastic ERP model 143, 1998.
using Simulated Annealing. To be
published. 28] Deb, K., and Goldberg D. E.; "mGA
in C: A Messy Genetic Algorithm in C.
21] Vujcic V. V. K., Cangalovic M. M.; IlliGAL Report 91008"; University of
"TABU Search methodology in Global Illinois at Urbana-Champaign, 1991.
Optimization"; Computers and
Mathematics with Applications, Vol. 37, 29] Lee K., and Horner A.; "Modelling
No. 5, pp. 125-133 Piano Tones with Group Synthesis";
Jour. of the Audio Eng. Soc., Vol. 47,
22] Battiti R., Tecchiolli G.; "Simulated No. 3, Mar. 1999.
Annealing and Tabu Search in the Long
Run: A comparison on QAP Tasks"; 30] Higuchi T., Kajihara N.; "Evolvable
Computers and Mathematics with Hardware Chips for Industrial
Applications, Vol. 28, No. 6, pp. 1-8, Applications"; Communications of the
1984. ACM, Vol. 42, No. 4, April 1999.

23] Paulli J., "Information Utilization in 30] Ko T. J., Cho D. W.; "Adaptive
Simulated Annealing and Tabu Search"; Optimization of Face Milling Operations
COAL Bulletin, No. 22, pp. 28-34, 1993. Using Neural Networks"; Jour. of
Manufacturing Science and Eng., Vol.
24] Chang C. S., Wen F. S.; "Tabu 120, May 1998, pp. 443-451
Search Based Approach to Trouble Call
Analysis"; IEE Proc: Generation, 31] Yalcinoz T., and Short M. J.;
Transmission and Distribution, Vol. 145, "Neural Networks Approach for Solving
No. 6, Nov. 1998. Economic Dispatch problem with
Transmission Capacity Constraints";
25] Franca P. M., Gendreau M., Laporte IEEE Trans. on Power Systems, Vol. 13,
G., Muller F. M.; "A Tabu Search No. 2, May 1998 pp. 307-313
Heuristic for the Multiprocessor
32] Veluswami A., Nakhla M. S., Zhang
Q.; "The Application of Neural
Networks to Em-based Simulation and
Optimization of Interconnects in High
Speed VLSI Circuits"; IEEE Trans. on
Microwave Theory and Techniques, Vol.
45, No 5., may 1997

33] Lu Y., Markward S. W.;


"Development and Application of an
Integrated neural System for an HDCL";
IEEE Trans. on Neural networks, Vol 6.
No. 6, Nov. 1997, pp. 1328-1337.

34] D.B. Fogel and J.W. Atmar,


Comparing genetic operators with
Gaussian mutations in simulated
evolutionary processes using linear
systems; Biological Cybernetics, Vol.
63, pp. 111-114, 1990.

35] Ingber L. A., Simulated annealing:


Practice versus theory; Journal of
Mathematical Computer Modelling;
Vol. 18, No. 11, pp. 29-57, 1993.
36] Siarry P., Berthiau G., et al;
Enhanced Simulated Annealing for
Globally Minimizing Functions of
Many-Continuous Variables; ACM
Transactions on Mathematical Software,
Vol. 23, No. 2, June 1997, pp. 209-228.

You might also like