Pareto Multi Objective Optimization

Patrick Ngatchou, Anahita Zarei and M. A. El-Sharkawi

Abstract-- The goal of this chapter is to give fundamental This simple optimization process is no longer acceptable for
knowledge on solving multi-objective optimization problems. The systems with multiple conflicting objectives. System
focus is on the intelligent metaheuristic approaches (evolutionary engineers may desire to know all possible optimization
algorithms or swarm-based techniques). The focus is on
techniques for efficient generation of the Pareto frontier. A solutions of all objectives simultaneously. In the business
general formulation of MO optimization is given in this chapter, world, it is known as a trade-off analysis. In the engineering
the Pareto optimality concepts introduced, and solution field, there are several examples of the need to perform trade-
approaches with examples of MO problems in the power systems off analysis. For example, designing distributed controllers
field are given. while reducing cost are two conflicting objectives. Similarly,
to place more functional blocks on a chip while minimizing
Index Terms -- Particle swarm optimization, Multi-objective that chip area and/or power dissipation are conflicting
optimization, Pareto optimization, trade-off analysis. objectives. To find the vehicle that covers the most distance
in a day while requiring the least energy is a multi-objective
problem. Minimizing the operating cost of a business while
I INTRODUCTION maintaining a stable work force is a conflicting objective
Optimization is an essential process in many business, optimization problem [[1]-[3]].
management, and engineering applications. In these fields, In power systems, the operation inherently requires MO
multiple and often conflicting objectives need to be satisfied. optimizations. For instance, in environmental/economic load
Solving such problems has traditionally consisted of dispatch, minimizing operation cost, minimizing fossil fuel
converting all objectives into a single objective (SO) function. emissions and minimizing system losses are some of the
The ultimate goal is to find the solution that minimizes or objectives that can be incorporated together to create a MO
maximizes this single objective while maintaining the physical problem. Minimizing the fuel cost, emission, and total power
constraints of the system or process. The optimization solution loss subject to stability constraints, generation capacity
results in a single value that reflects a compromise between all constraints, and security constraints requires MO
objectives. The art in this process is to formulate the function optimizations techniques [[4]]. Also, the optimization of
to achieve this desired compromise. reactive resources location and sizing of the transmission and
Conversion of the multiple objectives into an SO function distribution system is another example of a MO problem.
is usually done by aggregating all objectives in a weighted Some other conflicting objectives in the transmission
function, or simply transforming all but one of the objectives networks are transmission loss, transmission capacity and
into constraints. This approach to solving multi objective voltage stability. Other objectives in distribution networks
(MO) optimization problems has several limitations: 1) it include the distribution loss, power factor, and voltage
requires a priori knowledge about the relative importance of stability [[5],[6]].
the objectives, and the limits on the objectives that are Compared to SO problems, MO problems are more
converted into constraints 2) the aggregated function leads to difficult to solve, because there is no unique solution; rather,
only one solution; 3) trade-offs between objectives cannot be there is a set of acceptable trade-off optimal solutions. This set
easily evaluated; and 4) the solution may not be attainable is called Pareto front. MO optimization is in fact considered as
unless the search space is convex. the analytical phase of the multi criteria decision making
(MCDM) process, and consists of determining all solutions to
the MO problem that are optimal in the Pareto sense [[7]]. The
________________________ preferred solution – the one most desirable to the designer or
The authors are with the University of Washington, Seattle, WA decision maker (DM) – is selected from the Pareto set.
98195 (e-mail: [email protected]). Generating the Pareto set has several advantages. The
Pareto set allows the DM to make an informed decision by
seeing a wide range of options since it contains the solutions
that are optimum from an “overall” standpoint; unlike SO
optimization that may ignore this trade-off viewpoint. From a

1-59975-028-7/05/$20.00 © 2005 ISAP. 84

systems engineer’s perspective, this feature is useful since it II.2 Generic Formulation of MO Optimization
provides better understanding of the system where all the The general MO problem requiring the optimization of N
consequences of a decision with respect to all the objectives objectives may be formulated as follows:
r r r r r r
can be explored [[8]].
The goal of this chapter is to give fundamental knowledge
[ ]T
Minimize y = F ( x ) = f1 ( x ), f 2 ( x ),..., f N ( x ) ,
on solving MO problems. The focus will be on the intelligent
r (1)
metaheuristic approaches (evolutionary algorithms or swarm- Subject to g j ( x ) ≤ 0 , j = 1, 2, …, M
based techniques). We will motivate the use of these
techniques as opposed to the traditional mathematical r
programming approaches. The former are more appropriate Where x = [ x1 , x2 ,K , x P ]T ∈ Ω
for generic MO problems, particularly when information about
preferences or priority of the objectives is not known in r
advance, and when we seek to present the DM with a wide y is the objective vector, the g j ’s represent the
range of alternatives solutions. Hence, our focus will be on r
techniques for efficient generation of the Pareto frontier. A constraints and x is a P-dimensional vector representing the
general formulation of MO optimization is given in this decision variables within a parameter space Ω . The space
chapter, the Pareto optimality concepts introduced, and spanned by the objective vectors is called the objective space.
solution approaches with examples of MO problems in the The subspace of the objective vectors that satisfies the
power systems field are given. constraints is called the feasible space.
The utopian solution is the solution that is optimal for all
II MULTIOBJECTIVE OPTIMIZATION objectives.
r r r r
The objective of MO optimization is to find the set of x 0* ∈ Ω : ∀ x ∈ Ω , f i ( x 0* ) ≤ f i ( x ),
(2)
acceptable solutions and present them to the DM, which will for i ∈ {1 ,2 ,K , N }
then choose among them. Additional constraints or criteria
specified either before or after the search by the DM can help
guide, refine or narrow the search, but we will look at the
For N=1, the MO problem is reduced to an SO problem. In
generic case where there is no a priori information from the
that case, the utopian solution is simply the global optimum. It
DM.
always exists, even if it cannot be found.
II.1 Simple Example For the more general case where N>1 however, the utopian
Consider the problem of determining the most efficient solution does not generally exists since the individual
transportation mode. Assume two criteria are used to objective functions { f i } are typically conflicting. Rather there
determine this efficiency: (a) distance covered in a day, and is a possibly uncountable set of solutions, the so-called non-
(b) energy used in the process. The following transportation dominated solutions (for which an objective can not be
modes are considered: walking, bicycling, riding a cow, car, improved without degrading at least another one) that
motorcycle, horse, airplane, rocket, balloon, boat and scooter. represent different compromises or trade-offs between the
1/Distanc objectives.
e
II.3 Pareto optimality Concepts
To compare candidate solutions to the MO problems, the
concepts of Pareto dominance and Pareto optimality are
commonly used. These concepts were originally introduced by
Francis Ysidro, and then generalized by Vilfredo Pareto [[3]].
A solution belongs to the Pareto set if there is no other
solution that can improve at least one of the objectives without
degradation any other objective.
r
Energy Formally, a decision vector u = [u1 , u 2 ,K , u P ]T is
said to Pareto-dominate the decision vector
Fig 1. Transportation mode example. r
v = [v1, v2 ,K , v P ]T , in a minimization context, if and
Common sense can be used to obtain all potential solutions. only if:
r r
For example, the average car will need more fuel than the ∀i ∈ {1,..., N }, f i (u ) ≤ f i (v ),
average motorcycle, but the car can travel longer distance. r r (3)
Similarly, given the same amount of food, it would be and ∃j ∈ {1,..., N } : f j (u ) < f j (v )
generally expected that a cow would cover a smaller distance
than a horse. Using this reasoning, we can generate Fig 1
In the context of MO optimization, Pareto dominance is

85
 r r approaches consist of converting the MO problem into an SO
used to compare and rank decision vectors: u dominates v
r r r r problem, which then can be solved using traditional scalar
in the Pareto sense means that F (u ) is better than F (v ) for all
optimization techniques. Because most of these approaches
objectives, and there is at least one objective function for
r r r assume a priori information from the DM (either ranking the
which F (u ) is strictly better than F (vr ) . objectives in order of importance, or indication of target
r
A solution a is said to be Pareto optimal if and only if optimal values), these techniques are geared toward finding a
there does not exist another solution that dominates it. In other unique solution, the one that best satisfies the criteria and
r additional information (preferences) provided by the DM.
words, solution a cannot be improved in one of the
objectives without adversely affecting at least one other The second class of techniques, even though proposed
r r early, only took rise in recent years thanks to the advances in
objective. The corresponding objective vector F (a ) is called a
computational power and the development of population-
Pareto dominant vector, or non-inferior or non-dominated
based metaheuristic algorithms. These techniques are geared
vector. The set of all Pareto optimal solutions is called the
towards direct determination of the Pareto frontier by
Pareto optimal set. The corresponding objective vectors are
optimizing all the objectives separately. Some of these
said to be on the Pareto front. It is generally impossible to
techniques make explicit use of Pareto ranking. These
come up with an analytical expression of the Pareto front.
techniques are advantageous for real-life problems,
Pareto Front particularly those appearing in the context of power systems
since they present to the DM all possible, or at least a wide
range of trade-offs between objectives. This allows them to
make an informed decision.
f2 III.1 Classical Methods
F The classical methods consist of converting the MO
better

problem into an SO problem by either aggregating the

objective functions or optimizing one objective and treating
the other as constraints. The SO problem can then be solved
using traditional scalar-valued optimization techniques. These
techniques are geared towards finding a single solution and
better are ideal for cases where preferential information about the
f1 objectives is known in advance. It is also possible, by
maximize f1 , f 2 modifying the aggregation parameters and solving the newly
created SO problem, to approximate the non-dominated front.

Pareto Front a. Weighted Aggregation

The weighted aggregation method is a special case of the
utility function method, which converts the MO problem into
f2 an SO problem by applying a function operator to the
F objective vector. This function is designed by the DM to
better

capture his preferences [[2]]. A simple and popular utility

function is a linear combination of the objectives.
N
Minimize Z = ∑ w j f j ( xr ) with w j ≥ 0 and w = 1 (4)
N

j =1
∑ j
j =1

better where the weights (wj’s) can, for example, indicate the relative
f1 importance the DM attaches to objective j and must be
minimize f1 , f 2 specified for each of the k objectives a priori. The optimum
values of wj’s cannot be determined within the SO
optimization process. The solution of this SO problem yields a
Fig 2. Illustration of Pareto front for a biobjective single result that is as good as the selection of the weights.
optimization problem. Without prior information, choosing the weights can be
problematic. In a variant of this method called dynamic
Fig 2 depicts a Pareto set for a two-objective minimization weighted aggregation (DWA) [[9]], the weights are
problem. Potential solutions that optimize f1 and f2 are shown incrementally changed. For each new combination of weights,
on the graph. the problem is solved, thus generating a new compromise
solution. While simple to implement, this method does not
III SOLUTION APPROACHES generally yield the non-dominating front, and also misses
There are several methods to solve MO problems. Classical concave portion of the frontier. Furthermore, it is difficult to
control diversity along the Pareto front [[2],[10]].

86
 b. Goal Programming guide the search towards the Pareto-optimal set, and how to
A variation of the above technique is the goal programming maintain a diverse population in order to prevent premature
(or goal attainment) which seeks to minimize deviation from convergence [[12]].
prespecified goals. (5) is a common formulation.
N
Minimize Z = ∑ w j f j ( xr ) − T j (5) Compute
Compute
j =1 Test objectives
Constraints
where Tj represents the target or goal set by the DM for the
jth objective function, and the wj’s now capture the priorities Unfeasible
[[2]]. As in the weighted aggregation approach, the main Solution
drawback is the need for a priori information (priorities and Compare
Solution
targets). and add to
PSO, EC
c. ε-Constraint Pareto
archive
This is a method designed to discover Pareto optimal
solutions based on optimization of one objective while
treating the other objectives as constraints bound by some Pareto front
allowable range ε i . The problem is repeatedly solved for
different values of ε i to generate the entire Pareto set. Fig 3. Pareto front generation using population-based
r r r
Minimize f k (x ) , x ∈Ω, Subject to fi (x) ≤ ε i and techniques
r
g j (x) ≤ 0 . i = 1, 2, …, N; i ≠ k Evolutionary computing emulates the biological evolution
j = 1, 2, …, M process. A population of individuals representing different
Repeat (1) for different values of εi . solutions is evolving to find the optimal solutions. The fittest
individuals are chosen, mutation and crossover operations
This is a relatively simple technique, yet it is
applied, thus yielding a new generation (offspring). These
computationally intensive. Furthermore the solutions found
methods include genetic algorithms (GA), evolutionary
are not necessarily globally non-dominated [[6],[10]].
algorithms (EA) and evolutionary strategies (ES) which only
d. Discussion on Classical Methods differ in the way the fitness selection, mutation and crossover
Most of the traditional methods attempt to ease the operations are performed. Evolutionary techniques have been
decision-making process by incorporate preferential successfully applied to all sort of SO optimization, especially
information from the DM, and are geared towards finding the those where the objective functions is not well-behaved (not
single best solution representing the best compromise given differentiable, discontinuous, and/or no analytical
the information from the DM. Such techniques can formulation) [[3],[12]-[14]]. They often appear for the
approximate the Pareto front by essentially repeating the embedded optimization step of the traditional techniques
solution process after modifying the aggregation parameters presented in the previous section. Confusion should be
(weights or target levels). The obtained front while locally avoided between these hybrid evolutionary algorithm-based
non-dominated is not necessarily globally non-dominated. techniques, and the ones geared toward determining the Pareto
Furthermore, the distribution of solutions along the front efficient solutions presented in this section.
depends on efficiency of SO optimization solver [[11]]. While a. Non-Pareto-Based Approach: Vector Evaluated
the techniques are relatively simple to implement, they mostly Genetic Algorithm (VEGA)
are inefficient, and sometimes sensitive to the shape of the
Schaffer’s vector-evaluated genetic algorithm
Pareto front.
(VEGA)[[15]] is a non-Pareto-based technique that differs
III.2 Intelligent Techniques from the conventional genetic algorithm only in the way in
In contrast to the aggregation-based techniques, some which the selection step is performed. At each generation, the
intelligent techniques are geared towards direct generation of population is divided into as many equal-size subgroups as
the Pareto front by simultaneously optimizing the individual there are objectives, and the fittest individuals for each
objectives. Computational advances and the development of objective functions are selected (Fig 4). Regular mutation and
population-based metaheuristic algorithms contributed to the crossover operations are then performed to obtain the next
rise of these methods in recent years. Population-based generation.
algorithms have the advantage of evaluating multiple potential
solutions in a single iteration (Fig 3). In addition, they offer
greater flexibility for the decision-maker, mainly in cases
where no a priori information is available as is the case for
most real-life MO problems. However, the challenge is how to

87
 f2 f2 f2

f1 f1 V
f1
Fig 4. Illustration of VEGA approach for a biobjective
minimization problem. Each of the figures shows the selected Fig 5. Illustration of fitness computation for MOGA in a
individuals according to each objectives. The original biobjective minimization problem. Fitness of a given
population size is 10, and 5 individuals are selected for each of individual is proportional to the number of individual it
the objectives. dominates. For example, U which dominates 5 other
individuals, has a higher fitness, and hence probability of
selection than V which only dominates 2 individuals.
The VEGA algorithm is easy to implement; however, it
suffers from the speciation problem (evolution of species that In the Niched Pareto Genetic Algorithm (NPGA) [[18]],
excel in one of the objectives). This causes the algorithm to instead of bilateral direct comparison, two individuals are
fail to generate compromise solutions (those that are not compared with respect to a comparison set (usually 10% of the
necessarily the best in one objective, but are optimal in the entire population). When one candidate is dominated by the
Pareto sense). In addition, the algorithm is susceptible to the set while the other is not, the latter is selected. If neither or
shape of the Pareto front. both the candidates are dominated, fitness sharing is used to
decide selection. NPGA introduces a new variable (size of the
b. Pareto-Based Approaches comparison set), but is computationally faster than the
EAs in this category explicitly use Pareto-ranking in order previous techniques, since the selection step is applied only to
to determine the probability of replication of an individual. a subset of the population.
The basic idea is to find the set of non-dominated individuals Strength Pareto Evolutionary Algorithm (SPEA) [[19]]
in the population. These are assigned the highest rank and uses an external archive to maintain the non-dominated
eliminated from further contention. The process is then solutions found during the evolution. Candidate solutions are
repeated with the remaining individuals until the entire compared to the archive. A MOGA-style fitness assignment is
population is ranked and assigned a fitness value. In applied: fitness of each member of the current population is
conjunction with Pareto-based fitness assignment, a niching computed according to the strengths of all external non-
mechanism is used to prevent the algorithm from converging dominated solutions that dominate it. A clustering technique is
to a single region of the Pareto front [[10]]. A popular niching applied to maintain diversity.
technique called sharing consists of regulating the density of
solutions in the hyperspace spanned by either the objective f2
vector or the decision vector. The schemes presented below C
essentially differ in the way the fitness value of an individual
is determined prior to the selection step of the EA. Sharing is
often used in the computation of the fitness value. Mutation
B
and crossover operations are then performed to get the next
generation of individuals.
A simple and efficient method is Multi Objective Genetic A
Algorithm (MOGA) [16]. The fitness value of an individual is f1
proportional to the number of other individuals it dominates
(Fig 5). Niching can be performed either in the objective
space or the decision space. Fig 6. Illustration of fitness computation for NSGA in a
Another version is the Non-dominated Sorting Genetic biobjective minimization problem. A layered classification
Algorithm (NSGA), which uses a layered classification technique is used whereby the population is incrementally
technique [[17]]. All non-dominated individuals are assigned sorted using Pareto dominance. Individuals in set A have the
the same fitness value and sharing is applied in the decision same fitness value, which is higher than the fitness of
variable space. The process is repeated for the remainder of individuals in set B, which in turn are superior to individuals
the population with a progressively lower fitness value in set C.
assigned to the non-dominated individuals.
More recently, swarm intelligence approaches have been

88
developed for MO problems [[20]-[22]]. In particular, in Multi geared to finding a global optimum. In [[25]] the authors use
Objective Particle Swarm Optimization (MOPSO), the global goal attainment in combination with simulated annealing for
best (towards which the particles flock while exploring the the embedded optimization step to solve the problem.
search space) changed after a specified number of PSO epochs Different evolutionary algorithms were investigated in [[26]]
to a heuristically selected point of the emerging non- for the reactive power planning problems, but the cost
dominated front. The selection method is designed to function used is the aggregated sum of operation and
emphasize regions of low density, thus at the same time investment costs.
maintaining diversity. The algorithm also features a mutation As mentioned earlier, the drawbacks of aggregation-based
operator and a dynamic grid-based Pareto front management and other traditional MO techniques are the fact that they
mechanism [[22]]. require good knowledge of the systems in order to
appropriately determine how to weight the different
c. Discussion on Modern Methods
objectives. Also, only one solution is generated, which is not
The metaheuristic techniques successfully address the always optimal in the Pareto sense. More importantly, for the
limitations of the classical approaches when generating the system engineer’s perspective, these methods provide limited
Pareto front. Because they allow concurrent exploration of flexibility, and are not readily portable in case the operating
different points of the Pareto front, they can generate multiple conditions or requirements change.
solutions in a single run. The optimization can be performed From the system-wide perspective, liberalization of energy
without a priori information about objectives relative markets induced tremendous changes in power systems
importance. These techniques can handle ill-posed problems planning and operation [[8]]. The Independent System
(with incommensurable or mixed-type objectives). They are Operator (ISO) in particular, whose role is to ensure security
not susceptible to the shape of the Pareto front. Their main of the power systems must make quick and informed decisions
drawback is performance degradation as the number of after taking into account a multiplicity of competing
objectives increases, since there does not exist objectives. Because it is necessary to consider all the
computationally efficient methods to perform Pareto ranking. consequences of a decision, a tool that generates all possible
Furthermore, they require additional parameters such as compromises allows performing trade-off analysis and gaining
sharing factor, or number of Pareto samples which need to be better understanding of the system.
tweaked At first, traditional techniques were used to attempt to
generate Pareto front. For example, in [[8]], the ε-constraint
IV APPLICATIONS method is applied to problems an ISO would face in a modern
The power systems field has always been ripe with energy market (minimizing generation cost and real power
problems of MO nature: Economic dispatch, maintenance losses with or without reserve constraints; maximizing the
scheduling, reactive resource allocation, etc. These problems vulnerable operation, while minimizing real power reserve
are complex because they are often large scale problems for cost and minimize area import). The recent-years development
which the parameter space explodes as the size of the system of metaheuristic techniques and their application to concurrent
increases. In addition, the objectives and constraints are solving of competing objectives addressed the shortcomings
typically incommensurable and of mixed-type. of traditional technique. For example, the economic dispatch
These complex problems were traditionally approached with environmental constraints which has become the
using aggregation-based techniques with simplified environmental/economic dispatch (EED) problem since the
assumptions to reduce the complexity of the problem. Because rise in environmental awareness allows proactive treatment of
of the non-linear nature of the search space, stochastic search emission minimization as one of the objectives when
techniques such as GA, PSO or simulated annealing, which performing dispatch. In [[11],[27]], it is proposed to use
are robust enough to deal with these complexities, have been variants of NSGA to solve the EED problem. In [[4]] a
used successfully for the underlying optimization step. comparative study of popular population-based MO
The problem of thermal power dispatch with cost vs. evolutionary algorithms (NPGA, NSGA and SPEA) is applied
emission level trade-off is considered in [[23]]. The approach to EED problems and compared to traditional optimization
taken is interactive optimization by using the ε-constraint techniques. The research in [[28]] tackles the generation
method to generate non-dominated solutions. The goal scheduling problem, determining the operation schedule of
attainment method in conjunction with a GA-based optimizer generating units in order to meet a time-varying load with a
is used in [[24]] for tackling a stochastic version of economic variety of constraints (reserve constraints, security constraints,
dispatch where cost is minimized along with expected values etc.) while minimizing emission levels. In [[29]], capacitor
of deviations from target power and heat generation levels. location in radial distribution networks using a Pareto-front
Another common MO problem in power systems is the geared population-based technique in combination with Tabu
reactive power or VAr planning: this is the problem of search is proposed. In [[30]], a GA-based Pareto-based
locating and sizing capacitive resources in the most approach similar to NPGA is applied to the full Reactive-
economical manner so as to power losses and enhance the power Compensation Planning (RCP) problem.
voltage throughout the network. Early approaches for solving
this problem used mathematical programming and were

89
 V CONCLUSION 2001 IEEE Power Engineering Society Summer Meeting,
Jul 15-19 2001. 2001. Vancouver, BC.
In this Chapter, fundamentals of MO optimization are
[12] Zitzler, E., Deb. K. and Thiele, L. Comparison of
presented. Solution approaches were categorized into two Multiobjective Evolutionary Algorithms: Empirical
groups: classical techniques and intelligent techniques. The Results. Evolutionary Computation, 8(2), pp. 149-172,
former consist in converting the MO problem into an SO 2000.
problem. While simpler to implement, they require a priori [13] Jones, D.F., Mirrazavi, S.K. and Tamiz, M., “Multi-
information about the objectives, information which may not objective meta-heuristics: An overview of the current
be available at the outset. Furthermore, the solution obtained state-of-the-art”, European Journal of Operational
is not always optimal. More importantly, they are not suitable Research, vol. 137, num. 1, Feb. 2002, pp. 1-9.
for trade-off analysis. [14] Eiben, A.E., and Smith, J.E. Introduction to Evolutionary
The complexity of modern systems, particularly in the Computing. Springer, Berlin, 2003.
power systems requires the generation of a set of acceptable [15] Schaffer, J.D. “Multiple Objective Optimization with
solutions (instead of a single solution) that would allow the Vector Evaluated Genetic Algorithms.” Proceedings of
operator to choose from. The intelligent MO techniques, the International Conference on Genetic Algorithms and
which are population-based metaheuristics (evolutionary or Their Applications, Pittsburgh, July 24-26, 1985, pp. 93-
swarm-based algorithms) allow the concurrent solving of the 100.
different objectives, and generation of the Pareto front. [16] Fonseca, C. and P.J. Fleming. “Genetic Algorithms for
Multiobjective Optimization: Formulation, Discussion,
Generalization”. Proceedings of the fifth International
VI REFERENCES Conference on Genetic Algorithms, San Mateo, CA,
[1] Stadler, W. Multicriteria Optimization in Engineering & 1993. pp. 416-423.
in the Sciences, Plenum Press, New York, 1988. [17] Srinivas, N. and K. Deb. “Multiobjective Function
[2] Tabucanon, M.T. Multiple Criteria Decision Making Optimization Using Nondominated Sorting Genetic
Industry, Elsevier Science Publishers, Amsterdam, 1988. Algorithms,” Evolutionary Computation, 2 (3), 1994, pp.
[3] Coello Coello, C.A., Van Veldhuizen, D.A., and Lamont, 221-248.
G.B. Evolutionary Algorithms for Solving Multi- [18] Horn, J., Nafpliotis, N. and D. E. Goldberg, “A Niched
Objective Problems. Kluwer Academic Publishers, New Pareto Genetic Algorithm for Multiobjective
York, 2002. Optimization,” Proceedings of the First IEEE Conference
[4] Abido, M.A. Environmental/economic power dispatch on Evolutionary Computation, IEEE World Congress on
using multiobjective evolutionary algorithms: a Computational Intelligence, Piscataway, NJ. Vol. 1, 1994,
comparative study. in 2003 IEEE Power Engineering pp. 82-87.
Society General Meeting, 13-17 July 2003. 2003. [19] Zitzler, E. and L. Thiele. “An Evolutionary Algorithm for
Toronto, Ont., Canada: IEEE. Multiobjective Optimization: The Strength Pareto
[5] Begovic, M., Radibratovic, B., and Lambert, F., “On Approach,” TIK Tech. Report No. 43, Swiss Federal
multiple Volt-VAR optimization in power systems”, Institute of Technology (ETH), 1998.
Proceedings of the 37th Hawaii International Conference [20] Kennedy, J. and R.C. Eberhart. Swarm Intelligence.
on System Sciences, 2004 Morgan Kauffman, 2001.
[6] Nangia, U., Jain N., and Wadhwa C., “Surrogate worth [21] Song, M.-P. and G.-C. Gu. (2004) Research on Particle
trade-off technique for multi-objective optimal power Swarm Optimization: A Review. Proceedings of the Third
flow”, IEE proceedings generation, Transmission and International Conference on Machine Learning and
Distribution, Vol. 144, n.6, November 1997, pp 769-771. Cybernetics, pp. 2216-2241.
[7] Bagchi, T.P. Multiobjective Scheduling by Genetic [22] Coello Coello, C.A., Pulido, G.T. and M.S. Lechuga.
Algorithms. Kluwer Academic Publisher,1999. (2004). Handling Multiple Objectives With Particle
[8] Berizzi, A., M. Innorta, and P. Marannino. Multiobjective Swarm Optimization. IEEE Transactions on Evolutionary
optimization techniques applied to modern power Computation, 8(3), pp.256-279.
systems. In 2001 IEEE Power Engineering Society [23] Dhillon, J.S. and D.P. Kothari, The surrogate worth trade-
Winter Meeting, Jan 28-Feb 1 2001. 2001. Columbus, OH off approach for multiobjective thermal power dispatch
[9] Jin, Y., Olhofer, M. and B. Sendhoff. (2001). Dynamic problem. Electric Power Systems Research, 2000. 56(2),
Weighted Aggregation for Evolutionary Multi-Objective pp. 103-10.
Optimization: Why Does It Work and How? Proc. [24] Chang, C.S. and W. Fu, Stochastic multiobjective
Genetic and Evolutionary Computation Conf. (GECCO generation dispatch of combined heat and power systems.
2001). IEE Proceedings-Generation, Transmission and
[10] Coello, C.A.C., “A comprehensive survey of Distribution, 1998. 145(5), pp. 583-91.
evolutionary-based multiobjective optimization [25] Chen, Y.L. and C.C. Liu, Multiobjective VAR planning
techniques”, Knowledge and Information Systems, vol. 1, using the goal-attainment method. IEE Proceedings:
iss. 3, Aug. 1999, pp.269-308. Generation, Transmission and Distribution, 1994. 141 (3),
[11] Abido, M.A. A new multiobjective evolutionary pp. 227-232.
algorithm for environmental/economic power dispatch. in [26] Lee, K.Y. and F.F. Yang, Optimal reactive power
planning using evolutionary algorithms: a comparative

90
 study for evolutionary programming, evolutionary systems. He has published over 190 papers and book chapters, and holds five
patents. He is the author of two textbooks on electric drives and electrical
strategy, genetic algorithm, and linear programming. energy.
Power Systems, IEEE Transactions on, 1998. 13(1): p.
101-108.
[27] Rughooputh, H.C.S. and R.T.F. Ah King.
Environmental/economic dispatch of thermal units using
an elitist multiobjective evolutionary algorithm. in 2003
IEEE International Conference on Industrial Technology,
ICIT - Proceedings, Dec 10-12 2003. 2003. Maribor,
Slovenia: Institute of Electrical and Electronics Engineers
Inc., Piscataway, NJ 08855-1331, United States.
[28] Srinivasan, D. and A. Tettamanzi, Heuristics-guided
evolutionary approach to multiobjective generation
scheduling. IEE Proceedings: Generation, Transmission
and Distribution, 1996. 143 (6), pp. 553-559.
[29] Pires, D.F., A.G. Martins, and C.H. Antunes, A
multiobjective model for VAR planning in radial
distribution networks based on tabu search. IEEE
Transactions on Power Systems, 2005. 20(2), pp. 1089-
94.
[30] Li, F., J. D. Pilgrim, Dabeedin, C., Chebbo, A. and
Aggarwal, R.K. Genetic algorithms for optimal reactive
power compensation on the national grid system. IEEE
Transactions on Power Systems, 2005. 20(1), pp. 493-
500.

Patrick N. Ngatchou received his

undergraduate degree in Industrial
Engineering with emphasis in electricity
and electronics from the Institute of
Industrial Engineering, Liege, Belgium in
1996, and a M.S. degree in electrical
engineering from Columbia University,
New York, NY in 1999. He is currently a
Ph.D. candidate in electrical engineering
at the University of Washington, Seattle,
WA working on intelligent techniques for estimation and optimization.
Mr. Ngatchou is a member of NSBE and IEEE. He was awarded a Kaiser
Fellowship by the University of Washington’s College of Engineering in
2002.

Anahita Zarei received her B.Sc.

degree in Electrical Engineering from
the University of Washington in 2001.
She is currently working on her Ph.D. in
the area of Intelligent Systems Modeling
for Dentoalveolar Improvements.

Mohamed A. El-Sharkawi is a Fellow

of the IEEE. He received his B.Sc. in
Electrical Engineering in 1971 from
Cairo High Institute of Technology,
Egypt. His Ph.D. in Electrical
Engineering was received from the
University of British Columbia in 1980.
In 1980 he joined the University of
Washington as a faculty member where
he is currently a Professor of Electrical
Engineering. He was the vice president of the IEEE neural networks society.
He is the founder and co-founder of several International Conferences
including the Application of Neural Networks to Power Systems (ANNPS),
and Intelligent Systems Applications to Power (ISAP). He is the co-editor of
several IEEE tutorial books on the applications of intelligent systems to power

91