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

net/publication/4017059

Particle swarm optimization-based approach for generator maintenance

scheduling

Conference Paper · May 2003

DOI: 10.1109/SIS.2003.1202263 · Source: IEEE Xplore

CITATIONS READS

55 280

2 authors, including:

D. Srinivasan
National University of Singapore
349 PUBLICATIONS 16,464 CITATIONS

SEE PROFILE

All content following this page was uploaded by D. Srinivasan on 29 September 2014.

The user has requested enhancement of the downloaded file.

 Particle Swarm Optimization-based Approach
for Generator Maintenance Scheduling
Chin Aik Koay and Dipti Srinivasan
Department of Electrical and Computer Engineering
National University of Singapore
10 Kent Ridge Crescent
Singapore I19260

Absfracf- This paper introduces a particle swarm The first square bracket represents the dimension and
optimization-based method for solving a multi-objective second bracket represent the index of the particle. Each
generator maintenance scheduling problem with many particle stores its personal best position (pbesl) in memory and
constraints. It is shown that particle swarm optimization- the current velocity ( v ) and position (presenf). The index of
based approach is effective in obtaining feasible schedules global hest particle (gbesf) in the population is shared with the
in a reasonable time. Actual data from a practical power rest of the particles. The use of inertia weight ( w ) has
system was used in this study and results were compared improved performance on many test problems 141.
against those from other evolutionary methods on the
same set of data. Recent research has attempted to combine the attributes of
This paper also introduces a novel concept for spawning evolutionary computation with PSO concept to solve many
and selection mechanism in a hybrid particle swarm real applications with many constraints and objectives.
algorithm. T h e results suggest that this hybrid model Reference [5] uses co-evolutionary PSO technique to solve a
converges to better solution faster than standard PSO constrained optimization problem. First, Lagrange multiplier
algorithm. It is envisaged that this hybrid approach can he is introduced to transform the optimization problem into a
easily implemented for similar optimization and min-max problem with the saddle-point solution. Next, two
scheduling problems to obtain better convergence. PSOs work simultaneously with one PSO finding the
minimum part and the other focus on the maximum part of the
Index Term- Generator Maintenance Schedule, Artificial problem. At any one time, one PSO serves as an environment
Intelligence, Genetic Algorithm, Evolutionary Strategy, Particle to the other just like in evolutionary computation. Reference
Swarm Optimization [6] combines the features of ES and PSO to solve real world
applications in Power systems and Opto-electronics. The
hybrid model has proven to be effective over classical PSO in
1. INTRODUCTION many test problems.

P. article swami optimization (PSO) [ I ] has been

successfully applied in many areas such as function
optimization, artificial neural network training, fuzzy
The use ofselection in evolutionary technique [7] and
adaptive behaviour of PSO [8] have been investigated in some
system control, and other areas where evolutionary techniques recent works. Angeline [7] investigates the effect of selection
such as Genetic Algorithms (GA) and Evolutionary Strategy in particle swarm with standard test functions. The result
(ES) can he applied. Similar to Genetic Algorithms [2][3], suggests that selection may provide some advantage over
Particle Swarm Algorithm is a population based optimization classical PSO for certain functions. Adaptive PSO [8]that can
tool, where the system is initialized with a population of automatically track variations in a dynamic system is
random solutions and the algorithm searches for optima by introduced. A dynamic system changes its state and hence the
updating generations. In PSO, the potential solutions, called optimum value may vary. The proposed method is that if there
particles, are "flown" through the problem space by following is no improvement ofgbesf particle for a certain number of
the current optimum particles. The velocily and positions of iterations, then re-randomizes a percentage of the search
the particles are updated using the following equations: particles with the hope of finding the new global optimum
value. One drawback is how to set the fixed-duration number
PSO velociry update to give satisfying response time of the system.
v[ I[ ] = w*v[ I[ ] + 2*rand()*(pbest[ I[ ] -present[ I[ I)
+ 2*rand()*(pbest[ ][gbest] - present[ I[ 1) (1) All variants o f P S 0 (51-[SIhave similar features like
evaluation of fitness, modification of the current population
PSO position updaie either through PSO update or mutation and selection to
present[ I[ ] =present[ I[ ] + v[ I[ ] remove poor candidate solutions. The main contrast of the
various techniques is the different implementation of
evolutionary operators with the classical PSO. However, all
IEEE
0-7803-7914-4/03B10.0~2~3 167
the hybrid techniques have similar algorithm structure and (3)
perhaps can be broadly classified as follows: where:
X = unit in operation for that week
Pseudo code for PSO + Evolutionary Technique Y = unit in maintenance
I . Initialize population with particles T = length of the maintenance planning schedule (week)
2. Calculate fitness for each particle P, = generator output (MW) of operation unit
3. Introduce some evolutionary techniques and a,,h,. ci = fuel cost coefficient
parameters V , = maintenance cost per week ($/week)
4. For each patlicle, update its own pbest value if got D = downtime (weeks)
improvement
5 . For each particle, update its position by moving Penally cost is added to the evaluation function if the
towards the gbest particle or its own pbest position schedule cannot meet the power demand or the crew and
6. Repeat step 2 if terminating condition not satisfied resource constraints. Detailed explanations of each constraint
and formulation of the objective function can be found in [ 131.
In this paper, the focus will be on the application of PSO
concept to solve a generator maintenance scheduling that has 111. PROBLEM REPRESENTATlON
various constraints and objectives. Its performance will he
compared with results obtained by the heuristic, CA and ES
methods. Further variation of PSO will be suggested and Proper care has to be taken in the initial random generation
discussed in the hybrid PSO model with spawning mechanism of the candidate solutions due lo the following constraints:
to provide some adaptive ability over classical PSO.
Each generator should be taken off for maintenance
in c o n s e c ~ t ~ vweeks
e according to ire downtime.
In each week, the number of generators that can be
11. GENERATOR MAINTENANCE SCHEDULING -
maintained is limited to three due to resources and
PROBLEM DESCRIPTION
crew constraints
Maintenance schedule [9]-[121 is a preventive outage A generator can only be taken off once within a
schedule for generating units in a power system within a week. Hence, there should be no repeated values
specific time horizon. Maintenance scheduling becomes a within a respective week of the schedule. Except for
complicated problem when the power system contains a the numbcr 0,where it represents no generator to be
number of generating units with different specifications, and maintained.
when numerous constraints have to be taken into
consideration to obtain a practical and feasible solution. After much consideration, a useful representation of the
candidate solution is in the form of two-dimensional matrix.
Generator maintenance scheduling is done for time The IOWS of the matrix represent the number of weeks in the
horizons of different durations. Short-term maintenance schedule while the columns represent the index of the
scheduling for one hour to one day ahead is important for day- generators to be taken off for maintenance.
to-day operations, unit commitment, and operation planning of
power generation facilities. Medium-term scheduling for one 1V. EVOLUTIONARY APPROACHES FOR SOLVING
day up to a year ahead is essential for resource management. MAINTENANCE SCHEDULING
Long-term scheduling of a year to two years ahead is Generator maintenance schedules are typically generated by
important for future planning. power plant engineers who devise the schedule based on their
experience and knowledge of the system.
The domain of this problem is based on a real power system
of two industrial parks located in Bintan and Batam in A hybrid Fuzzy-genetic algorithm system was developed in
Indonesia. In the simulation, a planning horizon of 25 weeks this research [ 131 to handle this complex problem for which,
is considered in the generators scheduling problem. In each at present, there exists no effective planning tool due to the
week, there can only he a maximum of 3 generators in presence of various soft constraints and uncertainties. In this
maintenance due lo crew and resource constraints. The hybrid of fuzzy knowledge based system and CA, a fuzzy
maintenance of a generator must he done in consecutive knowledge based system evaluates the downtime of
weeks and care must be taken to ensure that the solution generating units, and several constraints. The fuzzy system
provides a feasible schedule. allows for uncertainties and impreciseness in evaluating the
number of operating hours, which give the flexibility in
Overall O&jectiveFunction IO minimize determining the downtimes. The knowledge of power plant
engineers is emulated in the form of if-then-rule to check the
The overall function to be minimized can he represented in a condition of the engines. The power of genetic algorithm is
compact form as follows: used to optimize the maintenance schedule giving the least
T X 2 Y T cost of maintenance and operation.
F= 2 X168(ax P, +b, P, + c , ) + Z VyD+ Z PcnallyCorl
1=lr=l y=t ,=l
168
 Another approach developed later replaced the genetic that PSO has combined these two concepts in one single
algorithm in this implementation with evolutionary strategy operation.
The results with ES were better than those with GA.
Crossover operation is implicitly implemented with the
Since evolutionary strategy emphasize 011 mutation as a gbestparticle sharing global information with the rest of the
search operator, a n algorithm for the mutation process was particles. Also. each particle can move towards itspbest
developed as shown in Figure I. Two heuristics were applied position in memory, hence crossover process within particle
to the standard ES algorithm in this problem, which results in itself can take place which may lead lo fastcr convergence.
better schedule with lower cost.
The implicit mutation process is represented by the
I. A new variable called multiple-mutations was introduced randomness introduced from the product of velocity v and
in the generation of the offspring solution matrix to model inertia weight w. For test functions optimization, w is often
the global search ability in the beginning and refined assumed to decrease linearly from 0.7 lo 0.4 lo model the
search towards the end. concept of global search ability in the beginning and refined
2. Since the selection is based on the best individuals, it is search towards the end of the run. [3]
highly possible that there may be repeated solutions with
the same fitness. In order to increase diversity and at the Hence, pure evolutionary crossover was modified to
same time not to discard other potential solutions. these implement PSO concept, which includes two important
repeated solutions will be discarded. In other words, all parameters: probability of crossover @cr) and probability of
the best individuals are distinct from each other to ensure mutation @mutate).
diversity in the next mutation process.
Whenpcr is set high, information crossover from
The results obtained with these two evolutionary gbest/pbest solution to current solution is encouraged. When
approaches are presented in Table I . pmutate is set high, random mutation within current solution
is encouraged rather than information crossover.

The update of the current solution is implemented with the

modified crossover operation that models PSO concept. The
current solution either moves towards gbest solution or its
ownpbest solution with equal probability.

A program to stimulate the modified crossover process in

PSO has to be carefully written to obtain feasible schedule
with the constraints met. The program starts with two input
schedules: gbesupbest and current schedules. For every
generator in the schedule, it either undergoes mutation process
with probability pmutate or crossover process with probability
(I-pmutate). For mutation process, the maintenance weeks
(positions) of that generator are randomly inserted in the
updated schedule (Figure 2). If crossover process is chosen,
then the schedule will be updated with information from
gbestipbest schedule with probability pcr (Figure 3).

In other words, ifpmutate (0.1) is set low, then most ofthe

time, the schedule will be updated with information from
gbestlpbest with probability per or current schedule with
Figure I : Mutation process in evolutionary slrslegv-bared algorithm for
maimlensnre scheduling probability (I-pcr).

V. PARTICLE SWARM OPTIMIZATION-BASED

APPROACH FOR MAINTENANCE SCHEDULING

A. Pure Particle Swarm Optimization-based Maintenance

Scheduling
From Equation I, one can infer that the particle will update
its velocity of flying by either moving towards the gbest
particle or moving towards its ownpbest position that it stored
in memory. Although PSO has no explicit crossover and
mutation search operation, but one can infer from Equation I
169
 - -
18 0
I I

14 0 D
1 14 0
1 0 0
1 0 0
1 16 U I ig o
16 0 0 16 0 0
. ... .. . .
- -
Canhdbte Same candidate
I
Figure 2: Grsphirll illustration oflhc mutation process where generator
"nil 1 i s randomly inserted in other conrerutire positions within lhe same
candidate solution
from the gbealpbert rchcdulc

Update ncwrchcduic
I

I I
Rcturnvpdated tchcdulc

Figure 3: Graphical i l l ~ ~ l r a l i oofthe

n C ~ O S S D Yprecess
~ ~ hctwccn two Figure 4 Flow Chart for PSO lmpkmenlstion
rindidales where P new candidate solution will be created and updated
with information either m m the first or second candidate
In the hybrid PSO model (SPSOES), an explicit mutation
This means that in the event when no random mutation takes parameter similar to ES was introduced. Each agent generates
place, ifgbest/pbest and current schedules are exactly the its off springs through mutation and the original agent is
same, then the current schedule will still remain the same even replaced if the off springs are better. This increases the
after update. Figure 4 shows the implementation of PSO to selection pressure within each agent as it has to compete with
solve the generator scheduling problem. its off springs. This approach greatly increases the flexibility
in programming as the population size is kept constant, but the
number of off springs for each agent can be varied easily with
B. Hybrid Spawning PSO with Evolutionary Strategv a single mutation parameter.
(SPSOES)
A novel hybrid approach that combines concepts from Fitness evaluation assigns fitness values to the candidate
particle swarm optimization and evolutionary strategy (ES) solutions based on the objective function to optimize. The ES
has been developed in this paper. From the results obtained mutation rate (mrate) determines the number of off springs to
for PSO and ES (Table I), it is evident that there is a lack of be generated from each parent. lfmrate is set to 3, then each
selection pressure in PSO. In ES, the 'survival of the fitness' parent will have 3 off springs. The variable mrate need not be
concept in the selection of the best individuals ensures that the a constant and can be set to any desired value during the run
population moves as a group towards better solutions. In PSO, and hence provides the freedom to increase or decrease search
each agent only compares his current fitness with the best ability to suit different applications.
fitness that it has stored in memory. This means that the
selection pressure is only contained within itself and not with Although mutation process dominates in ES, a small
other agents. probability of crossover (0.1) is introduced. This means that
90% of the off springs are generated through mutation of the
parent whereas 10% o f them are created through crossover
between the parent and a randomly selected parent in the
population.
 optima solutions, a check is performed to reject repeated pbest
solutions.

The update of the current solution is exactly the same as

pure PSO explained in section V (part A) shown in Figure 4.

VI. EXPERIEMENTAL RESULTS

Set crossover parameters
The maintenance schedule in the thermal system used for
this study is currently prepared based on experience and
considerations of the power plant maintenance engineers. The
engineer may take days or even weeks in scheduling the
maintenance of generating units, yet the schedule they made
may not be an optimal one. Moreover, the schedule does not
utilize the resources optimally. Furthermore, the engineers
Replace p m n t ifbetter currently take approximately one week to prepare the
maintenance schedule for 19 generating units for six month
planning horizon.
Spawn & e s t solution
Kcplace &en if better The following table compares this heuristic schedule
prepared by the engineers in a week, with the best schedule
obtained using the four approaches, namely, Genetic
Algorithm, Evolutionary Strategy, standard PSO and Hybrid
SPSOES. The results with GA were obtained with a
maximum iteration of400 with population size of 100. For a
fair comparison, the parameters for ES are set to be the same
as GA with a mutation rate of 1.

For both PSO and SPSOES, a maximum iteration of 1000,

Figure 5: Flow Chart for SPSOES model 1500 and 2000 with population sizes of 20,30 and 40
respectively are tested. To keep computation low, the ES
mutation rate in SPSOES model was set to 1 in the generation
For each parent, off springs are generated and they replace of off springs.
the original parent if they are better based on fitness values.
The schedules obtained by PSO and ES prove to be
An additional feature called the spawning mechanism was superior over GA and heuristic methods with much lower
introduced in the algorithm to introduce some adaptive ability cost.
into PSO. The concept was analogous to the natural adaptation
of amoeba with the environment. When amoeba receives The average cost of ES-based solution is slightly lower than
positive feedback, it reproduces by releasing more spores and standard PSO probably due to the selection pressure in ES.
hence increases the search ability for food. Where food is However, PSO is able to provide near optimal solution in the
scarce, it remains unchanged. Similarly, in potentially optima
shonest time possible with only a population size of 20.
region, it is encouraged to increase the number of search
Except for the case (Gen=2000, Parents=40) in PSO, ES
agents in the region. The amoeba concept was used to spawn
potential solutions found during the run. performs slightly better than PSO in terms of optimized cost
and worst case so far due to survival of the fitness concept.
To keep computation low, the spawning mechanism only With increased search ability in PSO (Gen=2000,
spawns the gbesr solution. Whenever a new gbest solution is Parents-40). PSO is able to obtain better optimized cost than
found, the number of spawns was set to IO initially and ES with the tradeoff for increase in timing.
decrement with each iteration till zero where the spawn The hybrid technique proposed in this paper, SPSOES
mechanism stops. If a better gbest solution was found during overcomes the limitations of PSO and evolutionary strategy
the spawn process, the number of spawns was reset to IO and by employing a synergistic combination of these two
the whole process repeats. approaches. The average cost of SPSOES-based solution, as
well as worst case so far, for the 3 tested settings is much
The surviving candidates are compared individually to its lower than standard ES or PSO. This means that ES or PSO is
personal best @best) fitness value so far stored in the memory likely to be stuck in a local optimum whereas SPSOES has a
and replace if better. However, to ensure diversity in the higher chance of escape from the local attraction. Although
171
the timings of SPSOES are higher when compared to PSO,
but SPSOES that
TABLE I: SILIULATION RESULTS

4w.

SPSOES
4457
--. .
~ ~~

4186- ~

- ...
,4965.

-
-
1 4 5 ~
2 i.
E 1m- -
. . . .
4 7 2 ~ : ,:
. ~,~ . ~ ~
,EsI ..................
...
incorporates spawning and selection mechanism provides the ............................ ......... ,..~
extra stochastic kick to get out of local optimum which results '9 ~ ......
in the most optimized cost so far.

Figure 4a compares the average convergence rate of the

fitness value between standard PSO and SPSOES for the case Figure 4s: Convergence rate of PSO and SPSOES
(Gen=1000, Parents=20). The last 500 out of 1000 iterations (&"=I 000,P.re"ts=10)
in the 10 runs are averaged and are plotted to give an
illustration of the improvement ofthe fitness during the run.
Hence, it can be observed that hybrid PSO with spawning of
the gbest solution generally converge faster to better solution
than standard PSO alone. Similar plots for the cases of
(Gen=1500, Parents=30) and (Gen=2000, Parents=40) are
given in Figure 4b and 4c respectively.

VI]. DISCUSSION

A. Hew-islic approach
Currently in practice, the maintenance schedule is obtained
by the power plant engineers on a trial and error basis. Due to
the many factors involved, scheduling becomes a complicated
problem. Heuristic method does not ensure optimal allocation .:
.. . . .". : . , . .,.
' Figure . .
of resources and is not versatile to changing environment 4b: Convergence rate o f PSO and SPSOES (C~n=1500,Parents=30)
factors. Hence, it is the most inefficient and has the highest
cost.
B. Generic Algorithm
CA attempts to add flexibility in the scheduling problem by
finding a set of optimum solutions through its population
based technique. GA emphasizes on the replacement o f . .
.... . ,.
individuals over time based on fitness. Those candidates with ..
higher fitness are more likely to be chosen as parents for the . I

next generation. Off springs are generated through cro~sover .-,

or recombination of the parents, replacing the original parents. ........ ..........
Occasionally, mutation may take place in the off springs.

The working principle of CA depends on the underlying

assumption that crossover increases thc genes pool where off
springs benefit from both of the surviving parents. In theory,
crossover may provide a means of escape from a local
optimum into other region and hence allows a more thorough Figure 6:Convergencerate o f PSO and SPSOES
search of the solution space. Although sound in theory, but in (Gen=ZOW,Parenlr=40)
practice, crossover between parents may not always produce a
better offspring. This probably explains why the result
112
 compared to GA or evolutionary strategy. The paper also
Unlike GAwhich replaces individuals. PSO models changes presents a hybrid spawning PSO and evolutionary strategy. In
in individuals over time and all individuals survive into the this approach, valuable features from both PSO and
next generation. The fate o f each individual is constantly evolutionary strategy are combined to provide a simple hybrid
altered based on the global optimal point discovered so far. model that is readily useable in many other applications.
This swarming effect, which cannot be found in GA, allows
the population to quickly converge into optimal regions ofthe Future work in this area will consider SPSOES with
search space. This probably explains why standard PSO can different ES mutation rates. Ways to improve on the timings
achieve better results than GA. will be considered. Further tests on SPSOES will be carried
out on a few benchmark test functions and performance will
be compared to the standard PSO.
C. Evolurionay S/ra/egis
The subtle difference between ES and GA is in the
parameter representation. ES works with real values of the REFERENCES
variables (phenotype) whereas CA works with binary strings
which are subsequently mapped to object variables. Since ES [ I ] Kennedy, J. and Ebcrflan. R. C, "Panicle swarm oplimizalion",
Proceedings of IEEE lnlemalional Conference on Neural Networks, vol. IV.
works completely on a phenotypic level, one can represent pp.1942-1948, Pisfataway. NI. 1995
more knowledge about the application domain into the coding
[2] Ebcrhan, Russell C. and Shi. Yuhui. "Comparison between genetic
o f the problem. Parents are mutated to generate off springs. In
algorithms and panicle s w a m oplimiralion" Evolutionary Programming VII:
the simulation, the best individuals are selected from the Proceedings afihe Seventh Annual Conference an Evolutionary
mutated and current population for the next population. This Programming, San Diego. CA. I998
ensures that the surviving individuals have a higher average [3] Lsvbjwg, M.. Rasmussen. T. K.. and Krink. T."Hybrid panicle swarm
fitness progressively. The competitive selection pressure opiimiser with breeding and subpopulations". Proceedings of the Genetic and
among candidates explains why ES can achieve a lower Evolutionary Compulalion Conference (GECCO 2001). 2001
average cost when compared to standard PSO.
(41 Eberhan. R. C. and Shi, Y. 'Comparing incnia weights and constriclion
factors in panicle swarm oplimizalion". Proceedings ofthe IEEE Congrrss on
Evolutionary Computalion (CEC 2W0). pp.84-88. San Diego. CA. 2000
D. Particle Swam Oprimizaliori Algorilhm
[SI Shi, Y. and Kmhling. R. A. "CO-evolutionary panicle s w a m oplimiwlion
PSO will be the most time efficient method to use when 10 solveMin-max problems", Pmcccdingr ofthe IEEE Congress on
looking for a near optimal solution as the population size Evolutionary Compuralion (CEC 2002), Honolulu. Hawaii USA, May 2002
can be kept small. The superiority of PSO over heuristic and
[6] Minnda, V. and Fonscca. N. "EPSO - best-of-two-worldn metu-heuristic
CA methods is clearly illustrated in Table 1 . However, the applied 10 power system problems". Proceedings o f the IEEE Congress on
average cost ofPS0 is higher than that of SPSOES. This Evoluiionaty Computation (CEC 2002). Honolulu. Hawaii USA. May 2002
implies that although PSO is able lo obtain satisfying
optimal solution within a short time span, it still lacks the [7] Angelinc. P. J, '"Using Selection lo improve panicle s w a m oplimization",
Proceedings ofihe IEEE Congress on Evoluiionaty Compuiaiion (CEC 1998).
ability to continuously improve upon the solution. Anchorage, Alaska, USA, 1998

[8] Hu, X.and Eberhan, R.C, "Adaplive panicle swarm oplimiralion:

E. Hybrid Spawned Parricle Swarm Oplimization and detection and response IO dynamic syslems", Proceedings oflhe IEEE
Evolulionary Slraregv (SPSOES) Congress on Evolutionary Compulation (CEC 2002). Honolulu. Hawaii USA,
May 2002
The hybrid approach proposed in this paper (SPSOES)
with spawning and selection mechanism proves to he [SI E. L. da Silva, M. Th. Schilling and M. C. Rafael. "Generalion
Maintenance Scheduling Considering Transmission Constrains", IEEE
superior over classical PSO in the cost obtained. Although
Transaclionson Power Syslem, Vol:15, NO. 2.2 May 2000.
SPSOES is not as time efficient as standard PSO, it provides
more consistent and reliable results. This can b e observed by [IO] Zura, H.H and V.H. Qiutahu. "Generationmaintenance schedding via
the low average cost obtained by SPSOES over the three successive approximation dynamic programming". IEEE Trms on PAS voI
experimental settings. The convergence graphs in Figures PAS-94, 1975
4 a - 4 ~illustrate that SPSOES is able to converge to better [ I I ] H. Sasaki, H. Choshi, Y. Takiuchi. J. Kubokawa. "A Solution of
solution faster than PSO. Maintenance Scheduling Covering Several Conseculive Years by Artificial
Neural Nerworks", Hiroshima Universily. IEEE 1993

[I21 K.P.Dahal, C.J: Aldridge, 1.R. Mc Donnald. "Cenentor Mainlenance

vI11. CONCLUSION Scheduling Using A Genetic Algorithm with A F u u y Evaluation Funclion",
Eselvicr, F u v y Sets and Systems 102: 21-29.1999.
This paper presents the results obtained from particle 1131 D Srinivasan D I M Malik. '.Flexible Generator Maintenance Scheduling
swarm optimization-based approaches to a practical In a Praclical System Using Fuzzy Logic and Genelic Algorithm", in Hybrid
scheduling and optimization problem for which evolutionaty Inrdlig~nenrSyarrms edited by A Abraham and M Kopponen. Heidelberg:
Spnnger Verlag, 2002.
computation-based approaches have previously been applied
and found to be effective. It is shown that particle swarm
optimization-based approaches yield superior performance
173

View publication stats