Proeedlngr of the 2004 IEEE

inbmatlonalConhnnrn on RoboUu 6 Automation

New means, IA April 2004

A Multiobjective Optimization Method for Strategic

Sourcing and Inventory Replenishment
Hongwei Ding, Lyks Benyoucef and Xiaolan Xie

MACS1 Project
INRIA (The French National Institutefor Research in Computer Science and Conlrol)
ISGMP Bat. A, Ile du Soulcy, Met2 57000, FRANCE
{ding & lyes.benyoucef & xie}@loria.fr
simulation-based multiobjective optimization
method is proposed in this paper for joint decision-making on
strategic sourcing and ioventory replenishment. More specifically,
a multiobjective genetic algorithm is developed to determine the
optimal supplier portfolio and inventory control parameters in
order to reach best compromise of two conflicting criteria: costs
and demand fill-rate. Discretwvent simulation is used to pronde
faithful evaluation of these two criteria. Numerical results on a
variant of a real case study are presented.

Abshoef-A

Keywords-ultiobjecdve
genetic olgorithm; simulofion;
supplier selection; inventory replenishment;

I.
INTRODUCTION
Supply chain design and management usually involve
multiple conflicting optimization objectives, such as low costs,
high quality, short leadtime and high demand fill-rate.
Traditionally, the total cost of all supply chain activities is used
as the key performance indicator (KPI) for supply chain
optimization.However, in the current competition environment,
it is not always desirable to reduce costs if this results in
degraded customer service level. Trade-off between these
conflictingperformance indicators should he made by decision
makers. Among numerous supply chain related problems,
strategic sourcing and inventory replenishment are two
research areas of significant practical impact. A large number
of research works can he found in the literature respectively
[1]-[8]. However, few of proposed techniques and methods
deal with the two problems at the same time. One of the
difficulties is the fact that strategic sourcing, principally
supplier selection, is a decision at the strategic level, while
inventory replenishment issues are rather at the operational
level. Strategic decisions have long-term impacts on overall
performances, which should be made in a comprehensive
manner. On the contrary, inventory replenishment emphasizes
short-term facets, which should respect to the dynamic feature
of demand. Existing optimization approaches for supplier
selection are mainlv based on deterministic mathematical

supply leadtime of suppliers have impacts on inventory

replenishment parameters. Reversely, diversified inventory
policies influence the requirements of supply leadtime. In such
an industrial context, a multiobjective optimization method is
indispensable in order to fmd best-compromised solutions,
which makes joint decisions in a comprehensive manner.
Genetic algorithm (CA) attracts more and more attentions
for various optimization problems. CA is a parallel and
evolutionaty search algorithm based on the Darwinian
evolution theory. It is used to search large, nonlinear solution
space where expert knowledge is lacking or difficult to encode
[lo]. In addition, it requires no gradient information, evolves
from one population to another and produces multiple optima
rather than single local one. These characteristics make CA a
well-suited tool for multiobjective optimization. Several
variants of multiobjective GA (MOGA) have been developed
in the literahue. MOGA is proving to he an increasingly
popular technique in solving realiitic industrial problems
[~21[161[m
In this paper, we propose a GA-based multiobjective
optimization method for joint decision-making on strategic
sourcing and inventory replenishment. More specifically, a
multiobjective GA is developed to perform stochastic search of
best-compromised solutions, which achieve trade-off between
costs and demand fill-rate. Decision variables, including
supplier portfolio, order assignment weight and inventory
replenishment parameters, are incorporated into discrete-event
simulation models for estimation of their impacts on
operational performances. The uniqueness of the proposed
method is that it not only makes decision at the strategic level,
but more importantly it addresses the operational aspects of
each solution through simulation. In summary, the
characteristics of this method are three-fold

.impact
- of scategic decisions on operational decisions ani do

not take into account important random phenomena. In reality,

the strategic sourcing and inventory decisions are closely
related and random phenomena such as demand volatility and
lead-time variability have significant impact on the
performances and costs of supply chain systems. Locations and

0-7803-8232-3/04/$17.00 02004 IEEE

2711

GA's evolutionary nature enables identification of

promising search directions of both sourcing and
inventory decision variables thanks to the performance
knowledge learnt f?om simulation
Simulation allows faithful evaluation of strategic
decisions, i.e. supplier portfolio, and the impacts of
uncertainties and risks on supply chain operational
performances, which is usually difficult for traditional
analytical optimization methods.

Multiobjective optimization enables practitioners

simultaneously handle both costs and other nonfmancial performance indicators, such as demand
fulfillment.

factors. A simulation optimization method was developed by [SI

to select suppliers &om operations point of view. Current and
Weber [7] showed that vendor selection problems might be
formulated within the constructsof facility location models.

The remainder of this paper is organized as follows. In

Section 11, the studied supply chain optimization problem is
presented. Related works on supplier selection, inventory
replenishment and multiobjective GA- are Summarized. In
Section 111, the proposed optimization method is described in
detail. Both optimization module and simulation module are
explained. A supply chain design problem derived &om a real
case study is presented in Section N.Numerical results are
analyzed and interpreted. Section V is a conclusion.

Motivated by risk and uncertainty reduction, more and

more enterprises pay attention to multiple sourcing, i.e. to
engage with more than one supplier at the same time. Order
splitting technique is often used in the presence of multiple
suppliers. A mixed integer programming approach is proposed
in [4] to decide simultaneously the set of suppliers and order
quantity allocations among them. Lau and Zhao [3] proposed
an optimal ordering policy for two suppliers. However no
inventory issues are addressed. Another common restrictive
assumption in preceding methods is that each replenishment
order is equally split among all the suppliers. On the other hand,
some papers took into account the ordering policy when
managing the inventory [1][2]. But they assume that products
of different suppliers are identical, i.e. the studied problem is
an "inventory-only" problem.

11.

PROBLEM SETTlh'G AND LITERATUREREVlEW

A. ' Problem Selfing

Supply chain optimization problems are generally
application specific.-Whatwe propose in the present study is
rather a problem-solving b e w o r k . In this paper, we consider
a supply chain consists of a set of candidate suppliers, a
distribution center and a customer. Fig. 1 illustrates the
network structure of the spdied supply cham.

Traditional mathematical programming methods have

difficulties in dealing with these two interrelated decisions
simultaneously. In fact, few of existing literatures studied it in
a comprehensive way. The interrelation of strategic sourcing
and inventory replenishment is missed in these approaches.

.The end customer generates random demands for a single

type of products. The distribution center holds stock to meet
C. Mulriobjecrive Oprimization Using Genetic Algorithm
demands and replenishes its inventory periodically. Assuming
Without loss of generality, a multiobjective optimization
that manufacturing is completely outsourced, all the
problem (MOP) can be expressed as:
replenishment orders are forwarded to the suppliers. Each
replenishment order is split into several sub-orders. The
quantity of each sub-order is determined p r ~ p ~ ~ t ito~the
~ l l y Minimize { f ; ( x ) , h ( x ),...,f,(x ) } ,subject tog(x)<O (1)
order assignment weight - of the target supplier. Suppliers
where x is the vector of design variables, A(.) is the ith
provide the same type of products but at different price, with
objective function and g(x) is the constraint vector. There does
leadtime etc.
not exist an absolute value. indicating the quality of solutions
for a MOP, in the absence of preference information. Solutions
The objective is to 6nd a set of solutions, which are best
are compared using the notion of Pareto dominance:A feasible
compromised regarding to several conflicting KF'Is, usually
solution x* is called Paretodptimal, if there exists no feasible
costs and customer service level. The decision variables consist
solution x such that:
of the supplier portfolio, order assignment weight lof each
selected supplier and inventory replenishment parameters.
forall i c { 1,...,in],
1;( x ) 2 1;(1')
(2)

B. Sourcing an8 Invenrory Replenishment

Selecting the most appropriate suppliers is considered as an
important strategic management decision that has long-term
impacts on enterprise performances. A number of methods
have k e n proposed in the literature, most of which fall into
three categories: cost-ratio method, categorical method and
weighted-point method [4]. Apart from them, mathematical
programming has been applied to perform quantitative analysis
on supplier selection problems [ 6 ] . The analytical hierarchy
process (AHP)is also used [5], where painvise comparison is
utilized to make trade-off between tangible and intangible

foratleast one i E { l ,...,m],

L(x)<L(x*)

(3)

Fig. 2 shows an example of optimization problem with two

objective functions where the Pareto fiont is marked with a
hold l i e .
Jones, Mirrazavi and Tamiz [9] presented an overview of
meta-heuristics methods utilized within the paradigm of
multiobjective programming. They stated that the classical
.. . ..
mininlize 11

Tx

..

feasible
-ion

'

"..

Figure 2. Pareto front of hvo objcctive functiOlu

Figure I. SrmCfure ofthe studied supply chain

2712

multiobjective programming techniques are more popular than

goal programming, interactive methods and compromise
prografming. In particular, 70% of the articles utilized GA as
the pmnary meta-heuristic in the last decade, in comparison
with simulated annealing and tabu search. The underlying
reason is the fact that GA can naturally propose multiple
solutions and therefore efficiently search Pareto-optimal
solutions.
As a pioneer in the field of MOGA, Schaffer [13] proposed
an extension of the simple GA (SGA), called Vector Evaluated
Genetic Algorithm (VEGA). This non-Pareto approach is
simple to implement, but in the case of a concave trade-off
surface, the population will tend to converge to a sub set of the
real Pareto-optimal fkontier, each of them particularly strong in
one of the objectives. Moreover, it is not able to produce
Pareto-optimal solutions in the presence of non-convex search
spaces. Fonseca and Fleming [I21 developed a Pareto-based
MOGA, combined with an external decision maker, to
determine a satisfactory solution for an engine design problem.
Each individual is ranked according to its degree of dominance.
This approach is relatively easy to implement, but its
performance is highly dependant on the niche size ash.
Srinivas and Deb [I41 adopted the ranking procedure suggested
by [IO] and proposed a nondominated sorting genetic algorithm
(NSGA). NSGA performs fitness sharing in the parameter
space to ensure a better distribution of solutions. However, it is
less effective than Fonseca and Fleming's MOGA and more
sensitive to the value of ab. Horn, Nafpliotis and Goldberg
[15] proposed a niched Pareto GA (NPGA) that does not use
ranking method. Rather, Pareto domination tournaments are
used to select individuals for crossover. NPGA runs very fast
but its performance depends on the value of tdOmFor more
details about GA basics and characteristics of different
MOGAs, reader is directed to [lo] and [ll].

nI. A MULTIOBJEC~VE
OPTIMIZATION
METHOD
The proposed method is composed of two modules: an
optimizer and a simulator. The optimization module is
developed using a simple GA [IO] with extensions for
multiobjective ranking, fitness sharing and Pareto-set filtering.
A simulation model generator is developed to produce discreteevent simuhtiou models automatically, regarding to various
solutions suggested by the optimizer. Both modules are
implemented using an objected-oriented language Ct+ to
achieve the modeling flexibility and computation efficiency. A
more detailed explanation of the two modules is given in
subsequent sections.
The flow diagram of the proposed method is shown in Fig.
3. For each individual in the present population, the optimizer
proposes a set of possible values for decision variables, more
specifically supplier poryblio, order assignmeni weight and
inventory control parameters. Then a primitive feasibility
check is performed before further steps. The repair procedure
repairs the supplier portfolio by adding more suppliers, if
current suppliers cannot provide all what the final customer
needs. With the verified set of parameters, the simulation
model generator creates a discrete-event simulator accordingly.
AAer simulation runs,the estimations of cosis and

..

Figure 3. Flow chaR of the proposed optimization method

demandfll-rare are retrieved and further used as the two

objectives to be optimized. AAer evaluation of the whole
population, a procedure assigns rank to each individual
according to its Pareto dominance degree at the present
generation. Corresponding fitness value is given depending on
the rank. In parallel;all the nondominated solutions in the
present population are reserved for further filtering. Like
simple GA, basic genetic operations, such as selection,
crossover and mutation are performed to produce the next
generation of solutions. These iterative procedures will
continue until certain stopping criterion is met, for instance the
preset maximum generation number is reached. Solutions in
the present Pareto-set provide decision makers a set of best-sofar compromised choices, based on which further studies and
analysis could be performed.
A. MOGA as the Optimization Module

Encoding of chromosomes is the first question to ask when

starting to solve a problem with GA. Fig. 4 illustrates the threesegment chromosome used in our MOGA. In the first part, a
string of binary genes is used to represent the supplier portfolio.
"1" means corresponding supplier is selected and "0" means it
is discarded. The second part represents the assignment weight
to each supplier, based on which replenishment orders are split.
The last segment comprises integers representing inventory
replenishment parameters. Due to their differences in nature,
genetic operations are performed in three segments separately.

2713

II

. . .

Binary A b i e s
(supplier selection)

. . .
-4

The customer is modeled as an entity where demands

are generated and material flows end. The demand
quantity follows certain random distribution or given
demand sample path. The demand interval could be
fixed or dynamic.

integer Variables
integer variable
(assignment weights) (replenishment)

Figure 4. Composition ofthe presented chrommome

Inventory replenishment activities are incorporated in a

building block named distribution center. The
distribution center is a facility where stocks of products
are hold to meet incoming demands. Various inventory
control policies are employed to make decisions on:
when and how much to replenish. Several classical
policies are implemented in this building block,
including base stock, R, Q , s, S.

Fitness is a value that reflects the quality of an individual.

The larger the fitness value of an individual, the higher its
chance of survival and reproduction. The essential difference
between a single objective GA and a MOGA is the method by
which fitness-is assigned to candidate solutions. For MOGA,
rank of an individual solution is generally determined by its
Pareto dominance in present population. In this study, we use
the ranking procedure devised by [12]. Consider, for example,
an individual i of generation I is dominated by p? individuals
in the current population. Its current position in the population
can be given by
rank(x,,t) = 1 + p,

(4)

All nondominated individuals are assigned rank 1, while

dominated individuals are ranked depending on their
dominance degrees. After the ranking procedure, fitness is
assigned accordingly [ 121.

In this study we treat the supplier selection problem as

an optimization problem. Several representative
attributes are integrated in the building block
supplier, including the product price, charged duties
for export and order-to-delivery leadtime etc. In
addition, a fixed investment cost is introduced
representing the cost for contract negotiation.

No niche techniques are implemented in the present

optimization module. For our problem, tbe most meaningful
and interesting range of the demand fill-rate is reasonably
between 80% and 100%. Too low a demand fill-rate is usually
not desirable for warehouse managers. The Pareto-optimal
solutions are not necessarily to he distributed on the fiont
evenly. A Pareto-set filtering procedure is introduced to
prevent loss of Pareto-optimal solutions. At each generation,
following the ranking procedure, all nondominated individuals
in the present population are copied and put into the Pareto set.
When new solutions are added into the set, a new round of
dominance check ,(filtering process) is performed and
dominated ones are discarded. Considering that the number of
Pareto-optimal solutions in OUT problem is limited, we did not
implement the replacement procedure, i.e. all found Paretooptimal solutions are retained:

B. Specifications of the Simulation Module

Discrete-event simulation (DES) is the most competitive
tool to model dynamic behaviors of supply chain. Its strength
in evaluating system variation and interdependencies makes
DES an appropriate technique for faithful evaluation of supply
chain performances in the presence of uncertainties and risks.
For most supply chain evaluation problems, various operations
are modeled in great detail to perform what-if analysis.
However, simulation plays a slightly different role in this
optimization method. In.the proposed method, DES is used to
evaluate the operational performances of candidate solutions.
Due to the random search nature of GA and the large solution
space, numerous runs of simulation are necessary. For the sake
of efficiency, a concise simulation model is a must. Hence, we
model the studied supply chain in an aggregated way.
Tbree building blocks are developed in C*, representing
the suppliers, the distribution^ center and the customer
respectively:

Transportation details are not modeled in this study. We

assume that when an order is sent out, the inventory is
replenished after a period of supply leadtime.
Due to the numerous combinations of decision variables, a
simulation model generator is indispensable. Unlike most
simulation optimization problems, in which decision variables
are only system control parameters, ow opfimization targets
incorporate the struchlre of supply chain. Different
combinations of candidate suppliers result in different supply
chain shctures. Correspondingly, information flows and
material flows in the simulation model are different fiom one
to another. Various simulation models have to be regenerated
each time. To overcome this difficulty, fmtly we implement
the three building blocks in an object-oriented manner to
constitute a supply chain modeling and simulation framework.
Based on the framework, we developed a simulation model
generator to produce different simulation models automatically
regarding to both supplier portfolio and inventory control
parameters given by the GA optimizer. More specifically,
given the supplier portfolio, the model generator initiates all
active supply chain facilities. When the supply chain structure
is built, parameters are set accordingly, i.e. order assignment
weight and inventory replenishmentparameters.

w. EXPENMENTAL
RESULTS
A variant of a real case study is presented in this section to
demonstrate how the proposed method can be applied to solve
realistic problems.
A. Method CusfomizationondAssumptions
We consider a supply chain composed of eight suppliers,
one distribution center (DC) and one customer. During the
simulation, the customer generates daily demands. The demand
quantity follows normal distribution witb a mean of 300 and a
variance of 50. All demands are collected by the distribution
center. The customer is defined as impatient, i.e. hackorder is
not allowed. A demand is considered as lost if there is no stock

2714

in the distribution center when it arrives. The distribution

center uses a weekly-review R, Q policy for inventory
replenishment, where R represents the reorder point and Q
denotes the order quantity. More specifically, a replenishment
order with a quantity of Q is generated if the inventory
position is shorter than the reorder point K. The two
inventory control parameters, R and Q , are decision
variables, suggested by the MOGA optimizer (see Fig. 4). Each
order incurs a fixed ordering cost, lOoE in the test case. The
daily unit inventory holding cost is set as 2%.
For a scenario with n suppliers, each order is divided into n
parts proportionally to the order assignment weight (an integer
between 1 and 32 in this study). Corresponding products will
be delivered to the distribution center after a period of supply
lead time. Table I provides the parameters concerning each
supplier. Considering the unique transportation link between
each supplier and the distribution center, the unit transportation
cost is included as a supplier characteristic.
For strategic sourcing decisions, especially in the presence
of randomness, simulation should run for a long period to
reflect supply chain dynamics. Thus, in this study, simulation
is set as 4 years. For each set of decision
horizon (T,j,,,)
variables, 4 replications are simulated using different random
number streams to smooth out residual randomness. After each
simulation nu& several KPIs are retrieved including: fixed
investment cost (CJ, purchasing cost (CJ, transportation cost
(C,), inventory holding cost (C,,), ordering cost (CJ, total
and lost demand (Qlm,). Fixed
demand quantity (eh.+)
investment cost is introduced when a specific supplier is
selected. Purchasing cost is calculated depending on price and
duties. Transportation cost is charged for each product. Then,
the two-fold optimization objectives are computed as follows:

Estimations of different KPIs are also listed. All these solutions

are composed of multiple suppliers, among which suppliers
Sl and S2 are always selected because of their lowest price.
We also discover that solutions 1, 2, 3 and 5 include the
supplier S8 into the supplier portfolio, which provides
products in the highest price but shortest delay. Consequently,
the lead time inventory is reduced due to the rapid
replenishment, while the purchasing cost increases
correspondingly. On the contrary, solution 4 chooses only two
suppliers Sl and 32. Lack of rapid inventory
replenishment results in a higher reorder point. As a result, the
inventory holding cost is higher than that of the others, while
the purchasing cost is lower for the exclusion of expensive
suppliers.
We note that most of the.solutions idthePGet0 fiont
comprise multiple suppliers and the supplier Sl is always
selected. For this reason, we design a local optimization
scenario, with only one supplier Sl, to justify the advantages
of using multiple suppliers. The simulation optimization
method is run, using the^ same. parameters setting, to find
optimal inventory parameters (R and Q). Fig. 6 highlights the
solutions with a demand fill-rate higher than 99.5% in both
optimization experiments. We observe that the local optimal
points, represented by + (results of the scenario with only
SI), are reasonably dominated by those found in the global
optimization, represented by 0. The relative difference
between the annual unit cost of the two best-so-far solutions,
with 100% demand fill-rate, is approximately 8.6%. The saving
rests in the reduction of lead time inventory holding cost, by
the introduction of more responsive suppliers, namely S2,
S6 and S8.
TABLE 1.

SUPPLIER CHARACTERISTICS

Minimize the annual unit costsh(x):

Maximize the demand fill-ratef,(x).

We run the MOGA optimizer for 2000 generations with a

population size of 100. The possibilities of crossover and
mutation are set as 0.9 and 0.1 respectively. Roulette-wheel
selection and one-point crossover are employed.

B. Numerical Results
Considering that customer service level is critical for most
practitioners, we focus on the analysis of solutions that keep
the demand fill-rate at a high level. Fig. 5 shows the
distribution of the best-so-far Pareto kont for solutions with a
demand fill-rate higher than 80%. These 49 points are evenly
distrihuted along the curve.
For further analysis, we select 5 solutions, which have the
highest demand fill-rates. Table 11 summaries the
corresponding supplier portfolio, order assignment ratio
(calculated by weight) and inventory replenishment parameters.

2715

80%

10

I1

12

13

I4

IS

16

17

Amdurdtcab(q

Figure 5 . Distribution of the best-so-farPareto set

18

TABLE II.

I4

IS

16
Amwl

SOLUTIONS

I7

I8

DETAILS AND VALUES OF SEVERAL KEY PERFORMANCE INDICATORS

19

ua coe (q

Figure 6. Performance cornparision for different optimization scenarios

V.

CONCLUSIONS

A simulation-based multiobjective optimization method is

developed in this study to make decisions on strategic sourcing
and inventory replenishment simultaneously. Discrete-event
simulation is employed for performance estimation. The
method enables practitionerj simultaneously optimize
decisions at both strategic and operational levels. Impacts of:
unceminties and riskson operational performances, which are
difficult for traditional analytical optimization methods, are
evaluated. For future research, we intend to implement other
variants of MOGAs and compare their performances on this
class of problem. On the other hand. some benchmarking,
works could be done by comparing the proposed method with
some analytical approaches for simplified cases.
ACKNOWLEDGMENT

This research lis partially supported by the European

Community research program, through the project ONE
(Optimization methodology of Networked Enterprises, Project
No. GRD1-2000-25710). These supports are gratefully
aclmowledged.
REFERENCES
[I] R. Ganeshan, Managing supply chain inventories: a multiple retailer,
one warehouse, multiple supplier model. Int. 1. of hod. Econ., vol. 59,
pp. 341-354.1999.

[2] R. Ganehan, T. Boom and A J. Stenger, The impact of inventory and

flow planning pmmetem on supply chain performance: an exploratory
study, Int. I. ofProd. Efon.,vol. 71,pp. 111-118, 2001.
[3] H. S. La and L. G. Zhao, Optimal ordering policies witb W Orupplierr
when lead timer and demands are all stochastic, European I. of
.
Operational Research, vol. 6 8 , ~120-133,1993.
[4] V. Jayarmas R Srivartava and W. C. Benton, Supplier selection and
order quantity allocation: a comprcbensivc model,The 1. of Supply
Chain Management, vol. 35, pp. 50-58, May 1999.
[5] S. H. Ghobypour and C. OBrien, A decision support system for
supplier selection wing an integrated analytic hierarchy process and
linearpro@&g,
Int. J. o f P r d . Econ., vol. 56, pp. 199-212, 1998.
[6] C. Weber and 1. R Current, A multi-objective approach U) vendm
sdectioa, Eumpean I. of Operational Research, vol. 68, pp. 173-184,
1993.
[7] I. R Cwent and C. Webcr. Application of facility location modelling
eonsmcts N vendor selection, European 1. of Operational Researcb,
vol. 76, pp. 387-392, 1994.
[a] H. Ding, L. Benyaucef and X.Xie, A simulation-optimizationappmacb
using genetic search for supplier selection, Proc. of tbe 2003 Winter
Simulation Conference, New Orleans, USA, 2003.
and M. Ta& Multi-abjective meta[9] D. F. loner, S. K M h z a ~
heuristics: an overview of the ament state-of-tbe-art, European I. of
Operational Research, vol. 137, pp. I-9,2002.
[IO] D. G o l h r g , Genetic algorithm in search, opthimion & machinelearning. Addison-Wesley, 1989.
[ I l l C. A. C. Coello, A0updated n w e y of ga-bared multiobjective
optimization techniques, ACM Computing Surveys, vol. 32, pp. 109143, lune 2000.
[I21 C. M. Fomeca and P. I. Fleming, Genetic algorithm for multiobjective
optimization: f m u l a t i o s disucrsion and generalktion, Race.of the
Fifth Int Cod. on Genetic Algorithms, S. Forrest, Ed Morgan
Kadmann, San Mateo, CA, pp. 416423, 1993.
[I31 I. D. Schaffer, Multipleobjective optimization with vector evaluated
genetic algorithms, hoc. of the Fin1 ht Cod. on Genetic Algorithms,
La-nce Erlbaum Associalea Inc., Hillsdale, NI, pp. 93-100, 1985.
[I41 N. Srinivas and 11, Deb, Multiobjeetive optimiration using
nondominated soning in genetic algorithms, Evolutionary Computation,
vol. 2, pp. 221-248, 1994.
[I51 I. Hom, N. Nafpliolis and D.E. Goldberg, A niched Parcto genetic
algorithm for multiobjective optimi2atiod: Roc. of the FLst IEEE Cod.
on Evolutionary Computation,pp. 82-87, Piscataway, NI, 1994.
[I61 1. A. Griffm, P. Schrodcr, A. I. Chipperfield and P. 1. Fleming, Multiobjeaive optimization approach U) the ALSTOM gasifier problem,
Roc.of the InstiNtioo of Mechanical Enginens, vol. 214, pp. 453-468,
2000.
[I71 F. Y. Cheng and D. Li, Genetic algorithm development for
multiobjective optimization of smcflues, AIAA Journal, vol. 36, pp.
1105-1112,1998.

2716