Sie sind auf Seite 1von 9

678 G E N E T I CA L G O R I T H M S

ANDCLASSIFIER
SYSTEMS

Evolutionary Algorithm For Structural Optimization

Mark S. Voss ChristopherNI. Foley


Civil & EnvironmentalEngineering Civil & Environmental
Engineering
MarquetteUniversity MarquetleUniversity
. I . 5 3 2 0 1 - 1 8I8
M i l w a u k e eW Milwaukee,WI 53201-188 I
mark.voss@marquette.edu 4 | 4 -288-6046 c.foley@ 41
marquette.edu414-288-51

Abstract Traditionalmultistorybuildingstendto havetheir


heaviestmembersnear the baseand their lightest
A hybrid rank-basedevolutionaryalgorithmthat membersnear the top. They also tend to graduallychange
takesadvantageof a-priori problem specific , the weight of their membcrsfrom floor to floor as one
informationand operateson a high.pardinality travelsup from the baseto the roof. [n this scnse,the
heuristic geneticrepresentation is presented
in buildingblockscould be seenas the structure'sbeams,
this paper.A rank basedfitnessstatement columnsand entirefloors.This suggests thatall
combinedwith generationally dependant penalty corresponding buildingcomponents(beams,columnsand
exponentsis proposedto condition the seven floors) and nearbybuilding colnponents(beamsand
componentsof the fitnessstatementso they columnsfrom nearbyfloors above/belowor an entire
participatefairly during the evolutionaryprocess. nearbyfloor above/belcw)could participatein
Translocationcrossoverand intclligent mutation meaningfulcrossoveroperations. This a-priori
wereutilizedto maintaingeneticdiversity.A knowledgeof the buildingblockscontainedwithin the
graphicalmethodis proposedto monitorthe multistorybuildingoptimizationproblemwas the
progressof the componentsof the fitnessfunction, motivationfor the explorationof a heuristictree
allowingthe userto interactwith the evolutionary representation for an individualand its designvariables.
process.Generationally dependantnon-linearrank
basedselectionwas usedto orchestratea soft Representation of designvariablesin a hierarchical
landingnearthe globaloptimumfor an example structuresuggestsreproductionwith crossoverof genetic
problemwith 20 discretedesignvariables. hierarchiesbetweenmatingindividualswherecrossover
occursat corcespondingor nearby locations. "The
easiestway to accomplishthis is to introducean
1 INTRODUCTION exceptionalcrossoveroperator,the translocation opcrator,
which producescrossing-over betweenrandomlychosen
non-homologous pairs" Holland (1975). Furthermore,the
The geneticalgorithmcan be classifiedas a stochastic
procedureand its successdependson the algorithm's extensionof crossoverto higher-orderrepresentations
(referredto here as macrocrossoveroperationson a
ability to effectively searchthe solution space,while
heuristictree representation)was anticipatedby Holland
exploitinggood solutionsthroughgeneticreproduction.
(1975)as a meanswith which to increasethe efficiency
The generalityof the classicalgeneticalgorithmis one of
its main assetssinceit can be appliedto a largerealmof of the geneticalgorithm.
problemswithout any a-priori problem specific
The ability to balanceexplorationof the solutionspace
information being required. However, this generalitycan
while exploitinggood solutionsis an extremelyimportant
causethe algorithm to spendtime searchingregionsof the
solutionspacethatare known to be unprofitable.A attributefor robustGA optimizationarchitectures.The
presentpaperseeksto illustratea generalevolutionary
questionthen arises: Giyen t*-o designsfor the same
structure,what are the strltcture's building blocksand in algorithm whereuponthe solution spaceis effectively
w'lnt meaninffil ways could they be exchangedin the exploredthrough probabilisticreproductionand fair
participationof penaltyfunctioncomponents throughout
crossoveroperationstypicallyfound itt genetic
the evolutionaryprocess. Furthermore,exploitationof
algorithms'.Holland (1975), Goldberg(1989),Holland
( I 9 9 6 ) .H o l l a n d( 1 9 9 8 t . good solutionsis performedthroughselectionpressltre
G E N E T I CA L G O R I T H M S
ANDCLASSIFIES
RY S T E M S 679

appliedat flexibie stagesduring the evolution of the whe-re: V is the volume of the cantilever,and I!, 6,, and
optimum-solution. Lastly, the paperseeksto presenta S, are the stiffness,deflection,and shapeconstraints
general,flexible GA architecturethat can be appliedto a respectively.
wide rangeof problemsoutsidethe simple cantilever
column exampleprovided. In order to employ a geneticalgorithm,the aboveproblem
statementneedsto be reformulatedas an unconstrained
optinization problem. Constraintfunctions for a
2 PROBLEM DESCRIPTION particularindividual are given below,

The rectangularcantilevercolumn shown in Figure I was o,;1 = r ( " i ' - r c . ' o:. = ^- A-o_a l - ^ -. , i

t tt
usedto study the effectivenessof the evolutionary
algorithmproposed.
zl o;, - a),- a'r",^ ooir , = u;, - 6;:"' (3)
I

ol ^t ^eoal
o; = s.j - srt"'

3 ALGORITHM FEATURES

The traditionalgeneticalgorithmworks well when one is


optimizingmuiti-variatefunctionswhen thereis no a-
priori knowledgeof the interrelationshipsbetweenthe
solutionvariablesof thepbjectivefunction. For example,
w h e no n ek n o w st h a tt h e f u n c t i o n , f ( a , b , c , d ) i s
m i n i m i z e dw h e n , a < b < c < d t h e t r a d i t i o n agl e n e t i c
algorithm is committedto searchingunprofitableregions
of the solutionuniverse
Figure 1: Segmental
CantileverColumn
One methodto remedythis situation is to modify the
penaltyfunctionsresidingin the fitnessstatement.Thus,
Two designvariables(h^ and hr) are possiblefor eachof a constraintfunction of the fbllowing forrn,
the ten segmentsand the valuesfor thesevariablesare
assumedto takeon discretequantities.The derivation @ u , d "=, ^ r u l @ b ) , 0 ] - ^ o x l { b c ) , 0 ] ,' '^' \
and statementof the magnifieddisplacementand stiffness , m c u l( r ' r 0 . o ]
equationscan be found in Vossand Foley (1999). could be used to movethe populationin the directionof
promisingregionsof the searchspace.
In the current study, the volume of the structureis
consideredas the characteristicquantityto be optimized. A-priori knowledgeof the design variableinterrelation
It follows that the optimizationproblemmay be expressed also suggestsan ordering of the classicalgenetic
as, algorithm'sbinarystringrepresentation, suchthat,
Minimize:
l0 e < b s c < d -goal
t lt,
- t,_,).h,.h, (l)
1 0 0 0l 101 0I | 1 0 0 0110 0l I
Subject to: This providesa basisfor hybrid macro crossover
--i
1(_ > K Soal
"T *', , rc\""' operationsbetweenindividuals whereby two individuals
selectedfor crossoverwould exchangethe value of whole
-toal ^ i
nearby variables(translocationcrossover). Although
6t'"' ' 6r, o; ro, translocationcrossoveroperationsmake intuitive sensein
(2) this case,they are not encompassedby the Schemata
Srtoo' - S, sr"'''' s'" Theorem;Goldberg(1989).

,U
fr"
, ,L
2 frr 2 ft, tt.u>tt>tt.! The presentstudy proposesa problem specific heuristic
)ft
Michalewicz (1992), of design
tree representation;
680 G E N E T I CA L G O R I T H M S
ANDCLASSIFIER
SYSTEMS

variablesas shown in Figure 2., The tree representation component(s;giuen a rank of one with the next largest
facilitatesthe recognitionof building blocks usedin given a rank of two, etc. Weight and constraint
exceptionalcrossoveroperationsinvolvinghomologous componentsthat have the samevalue are assignedthe
and non-homologouspairs and could also be thoughtof samerank. Once the ranks have beenassignedthey are
as a geneticprogram with a static representation; multiplied by a scalingmultiplier and addedto a constant.
Banzhaf,et al. (1998),Koza (1996). Nodesa, b, and c The result is then raisedto a GenerationallyDependant
representlocationsfor x-dimension,y-dimensionand Penalty Exponent (GDPE) to tune the relative weight of
whole level crossover,respectively.The nodesare eachcomponentin the objectivefunction. For example.
labeledat the secondlevel, but shouldbe consideredas given the following numericalvaluesfor the
generalizedlocationsfor crossoverat a given hierarchy displacementconstraintviolation for an assumed
level. populationof sevenindividuals:
lst LEWL
hr,
hv, Oa= 6t-68o4/
hr, = { r : , - s 2 , 2 s , 2 ,- l o o o, 3 4 2 ,1 3 \
2nd LE/EL
hy,
the ranks of the displacementconstraintviolations for this
populationare then defined as,

gth LanEL hx- R ( ( D 6) = { Z , O ' Z ' r , o ' 4 ' 2 \


nv"
hrro The rank of the displacementconstraintviolation for the
lmh LA1EL
h yro sixth individual in the populationis denoted,
l-igure 2: HeuristicRepresentation R6(O6) = 4

The principal of minimal alphabets(which is supported The componentpenalty with respectto displacementfor
by the SchemataTheorem),advocateslow order genctic the i'h individualin the population,isdefinedas,
representations (low cardinality)for maximum
f6 = [l ' E.R(o.)l'o
effectiveness;Coldberg( 1989). Translocation crossover l.orol

increases diversityover generations by allowing


emigrationof variablesfrom one genelocationto another. where:{6 is the multiplierfor the deflectionconstraint.
This emigrationincreasesthe diversity of potentialvalues Similar valuesare definedfor volume,stiffness,and
that a genecan expresswith a given genetic shape. ln the presentformulation, all valuesof ( are 1.0
representation. If this diversityis maintained,
the as indicatedin Table l. For exarnple.the displacemerrt
principal of minimal alphabetsmay be relaxed;Mitchell penaltyfor the sixth individualin the populationis,
(lee6).
/rt=tl*(u.41'o

4 RANK BASED FITNESS The rank basedfitnessfor an individual ten segment


cantilevercolumn can then be written as follows;
Thereis a subjectivecomponentto optimizationin that a F, =.f,* * f,*'^ * f,*" * t!' t f!' r5)
designermight be willing to live with a few lessthan
* J,.5r + J,.Sy
optimal componentsto get a desiredresult. Therefore,a
highlydesirablealgorithmwould containoptions
allowing the userto imposethe importanceof indi.iidual Table 1: User DefinedConstants
components. This, in turn, would give control to a Medium
seeminglyrandomoptimizationprocedure.The above No Shape Shape l-arge Shape
considerationsmotivatedthe developmentof a rank based Constant Penalty Penalty Penalty

optimizationstatement.The rank basedoptimization a 0.000 0.500 1.000


statementproposedis easily implementedand somewhat
semi-automatic. The designermustassigna scalar 0.000 2.000 2.000
multipleand exponentto eachoptimizationstatement r.000 1.000
1.000
componentbasedon both objective knowledgeand
subjectivepreference. Yr.Yr 0.030 0.030 0.030

\z 0.003 0.003 0.003


The weight and all constraintcomponentsof each
individualare rankedwith ali components lessthanzero C ' . Q( ' 0.190 0.190 0 . 19 0
given a rank of zero and the smallestnon-negative
1,,.1", ,,. 3.000 3.000 3.000
G E N E T I CA L G O R I T H M A
SN D C L A S S I F I E R
SYSTEMS ool

5 ALGORITHM COMPONENTS probability density function to determinethe rank of the


individualto be selected.
'1.0
GenerationallvDenendentNonlinear Rank Based
0.9
Selection;naik (f^qqO),Michalewicz(1992), combined
0.6
with GDPE(s) was employedto allow dynamiccontrol of
C 0.7
algorithmconvergence.By ranking the rank basedfitness
I 0.6
valuesthemselves(eachindividual given a fitnessrank
€€ o.s
from 1 to the populationsize) and then using GDNLRBS,
it was possibleto dynamicallycontrol how fast the 95 o'r
E * o.g
selectionpressurewas increasedfrom somewhat 5
E 0.2
egalitarian(low) pressureduring early generations,to 5
() o.t
high (focusedon the exploitationof fit individuals)
0.0
during later generations.It should be notedthat o 10 20 tor,r.T" 60 70 8o eo
GDNLRBS is a global selectiontuning mechanism *"1
whereasthe GDPE(s) are local selectiontuning Figure 3: GenerationallyDependantNonlinear Rank
mechanisms.The interrelationof thesetuning BasedSelection
mechanismsis integralto the efficient operationof the
proposedalgorithm. Elitism was alsoemployedin the proposedevolutionary
algorithm.The top two individualsof every generation
The generationallydepencientpenalty.*pon.n,, o." are carried into the next gencration. Since GDNLRBS is
definedby, used,the algorithm has a lessenedtendencyto exploit
super-fit individualsthan proportionalfitness selection.
lc' u "
n I. - - a + 'gt 'Gl (6) This allows the benefitsof elitism to be employedwithout
Lm6
prematureconvergenceon super-fit individuals during
early generations.
where: a and p are userselectedconstants;G.,,, is an
integerrepresentingthe current generation;and G.* is an Figure 4 illustratesthe flow diagramfor the proposed
integerdefining the maximum numberof generationsto hybrid geneticalgorithm.
be carriedout in the geneticalgorithm.By choosingthe
valuesof c and B carefullyit is possibleto focusthe
evolutionarysearchon different penaltiesat different
points during the evolution of the optimum design.

The generalform of the selectionprobability for Rank ComFnents of Fitness Funclion


s6KW
crossover,cary-over, and mutation is given by:
q(r q)' .'
rp\('rl)r = - ,- (1)
(l-s)']
Rank Individual
Fltness Values

where:r is the rank of the individual;ru is the population


Mating - Non-Llnetr Rank-Brsed Selectlon
size;and q e (0. .l). Larger valuesof 4 imply stronger
selectivepressureof the algorithm; Michalewicz (1992).
A generationallydependentselectionpressureparameter,
q, is defined as,

q=y*( ,
t'
Grur, * 5
(8) 23.4 % Parl Seg Homologous
w
2.6 o/oParl Seg Tran3l@ation
G^ 2.0 % P.d Seg TEnslocatlon - Scll
]'
where; y, I and ( areuser defined constants.All 13.6'lo Full Ssg Hohologour
1.8'l" Full Sqg Translocatlon
4.0 o/oFull Sog Tranalocation - Selt
multipliers,constantsand exponentsusedin the proposed
algorithm are given in Table l

A cumulativeprobability density function (Figure 3) is


constructedusing the nonlinearprobability function,p(r).
Figure 4: EvolutionaryAlgorithm Flow Diagram.
An individual can then be chosenby selectinga random
real numberr € (0...1)and mappingit onto the cumulative SeparateGDNLRBS functions were used for selecting
individualsfor participationin crossover.carryoverand
682 G E N E T I CA L G O R I T H M A
SN D C L A S S I F I E R
SYSTEMS

intelligentmutation. It was felt that creationof separate By simply tuningthe convergence paramcters. it is
GDNLRBS distributionsfor crossover((, , y, . l,) , possibleto orchestratea soft landing close to the global
carryover(C,, \ z, 1,,),and mutation ((: , y: , l.) did not optimum.
overly complicatethe algorithm and allows for maximum
flexibility with respectto tuning.The threecomponents
are usedin conjunctionwith one anotherto tune the 6 RESULTS AND DISCUSSION
evolutionaryalgorithm. Percentagesfor reproduction
usingGDNLRBS arealso given in Figure4. The cantileverexampleproblem developedwas given a
fixed lengthof 500 inchesdivided into l0 equalsegments
In traditionalcrossoveroperationsit is possibleto define of 50 inches. The orientationo[ the loadsand degreesof
the amountof materialswappedduring reproduction, freedomare shown graphicallyin Figure 1. Each of the
sincethereis generallyonly one type of crossoverused segmentscould vary between10 and 80 inchesin both the
(uniform, single/doublepoint, etc.). The presentstudy x- and y-directionsadheringto discreteincrementsof
implemented6 exceptionalcrossoveroperationsduring 0.35 inches. An axial loadof P. = 80,000kips combined
reproduction.Due to the complexityof thesecrossover with horizontalloadsof H^ - Hy = 50 kips were applied
mechanismscascadingover one another, it is not at the top of the cantilever. The proposedevolutionary
possibleto directly set the amount of geneticmaterialthat algorithmwas run for 50 generations with a fixed
is swappedduring reproduction. It was therefore populationsize of 80. Table I should be referencedfor
necessary to run a simulationto determinethe percentage informationregardingthe evolutionaryalgorithm
of geneticmaterialexchanged;Voss and Foley(1999). parameters usedfor all runs. The constraintsassignedfor
The averageamountof geneticmaterialswappedbetween the problemaregiven in Table 2 below.
individualsand swappedinternallyduringreproduction
was found to be 4l .57oand 67orespectively. This resuLts Table 2: ConstraintParanreters
in a total average reproductit'e genetic modif cutionfront
crossoverof 17.5 Va. Constraint
Quantity (Goal) Magnitude
The crossoveroperations(operatingon generalizednodal *r_,,*r,
locationsa, b, and c) are listedand discussed as follows:
0,0
l.\ Homologous Parti.al Segment Crossover:
corresponding "a" or "b" locationson two unique 6, ,6, 2 . 0, 2 0 . 0
individualsarecrossedover.
2.) N on-homologous (Translocation) Partial Segm ent s,,s, (4)
0,0: referto equation
Crossover:"a" or "b" location crodsedover with an
"a" or "b" locationoffset up to four levelsaway on ,U,L
fr, , fl, r 0 ,8 0
two uniqueindividuals.
3.) N on- h om o Iog ou s (T ran sI o catio n ) Self Partial
,U,L
SegmentCrossover:"a" or "b" locationcrossed fi- , ff,, 1 0 ,8 0
over with an "a" or "b" location offset up to four
levelsaway on the sameindividual.
Figures5, 6 and 7 showthe componentrankingof
4.) Homologous SegmentCrossover:Corresponding
fitnessfunctioncomponentsfor individualsin the
"c" locationson two uniqueindividualsarecrossed
populationat various stagesin the evolutionaryprocess
over.
with varying GDPE magnitudes.The graphsare
5.) Non-hontologous (Translocation) Segment
constructedby plotting the sortedvaluesof the rank based
Crossover:"c" locationcrossedover with another
componentpenaltyfunctions. It should be noted that the
"c" location offset up to four levels away on two
independentvaluesdo not necessarilycorrespondto the
uniqueindividuals.
sameindividualfor all componentplots. The plots
6.) Non-homologous (Translocation) Self Segment
illustratethe relativecontributiona particularcomponent
Crossover:"c" location crossedover with another
playsin establishing the fitnessof individualsthroughout
"c" locationoffsetup to four levelsawayon the
the populationduringthe evolution. It is easyto monitor
sameindividuals.
thesegraphsto determineif any one componentis
Figure 4 illustratesthe percentagesof eachcrossover
dominatingthe selectionprocess.The usercan easily
operationin the proposedalgorithm.
modify the parametervaluesassociatedwith a dominating
componentto reduceits contribution throughoutthe
The proposedalgorithm doesnot employ a criteria for
population. The effectsof any modificationsare then
termination. Instead,the algorithm usesthe GDNLRBS
observedthrough the useof theseplots.
to force convergenceafter a given numberof generations.
ooJ
SYSTEMS
GENETICALGORITHMSAND CLASSIFIER

in the
90 trw Figuie 5 illustratesthat only a few individuals
both deflection and stiffness
r 6, ini-tirt popututionare without
89 Since the value of a
violationi in both directions'
o6 a constralnt
70 componentrank for an individual without
^K to that component is zero' the
60 violation with respect
aK is apparent in Figure 6' In this
-50 prog"tt of the algorithm
t s, that only 24 individuals (of the
:. 40 fu;", it can be seen
- have constrai:rt violations with
'-+ 9S
v
pJpulation 80 total)
30 components'
i"rp"., to deflectionand stiffnesspenalty

is evidentthat the
ContrastingFigures6 with Figure 7 it
the convergence
large shapepenaltyexponentis delaying
30 40 50 60 70 and stiffnesscomponents' In Figure 6'
10 20 of itr" a"h".rion
lndividual (sorted) to dominatethe
the shapecomponentsare not so large as
- Generation penalties.areapplied
Figure 5: Ranking of FitnessComponents ,"t".tion pro."rr' When large shape
the
Number l, Medium ShaPe GDPE' as in Figure 7, the shapecomponentsare dominating
selectionProcess.
90 for the
oW Figure 8 is a plot of the algorithm converge.nce
objectiveof the
80 oD beit individual of eachgeneration' The
volume
70 ob optimizationprocedurewas to find the minimun'r
Figure 8 is
cantileuer meetingall constraints' Therefore'
v
60 K lo versus the
given as a plot of ihe generationnumber
rO-
dv&
- ls)^ of that
^ 50
H
A K 6 iolurn" (in cubic inches),ofthe best individual
lf,)^
-c
1F
40 s A,A
404
lo^
generation.
&l
30 S &Lz
15 0 0
20
10
13 5 0
0 9(
10 20 30 40 50 60 70 80
0
o
lndividual (sorted) 12 0 0

- Generation
Figure 6: Rankingof FitnessComponents E
ShaPeGDPE' 10 5 0
Number20' Medium

160 oW

140 . 6,
5 l0 15 20 25 30
o 6v
120 Generation
^ K,
- 100 Trajectories- BestIndividual
K., Figure 8: Convergence
-,.
80 s, the
v
One needsto be careful when interpreting
60 S. plots. The most meaningful information can
*nu".g"n..
plot' lf
40 U" inf#.a from the trajectoryof the convergence
of a
very small penaltieswere applied at the beginning
20 *n, tt uoiurn. would be dominantin the selection
" would
process.In this case, the convergenceplots
below' As the
10 20 30 40 50 60 70 80 Ipp."""ft the global minimum value from
to a larger
Individual (sorted) .trpon.n, penaltiesbegin to participate
than the
- Generatton .^t.nt, heauierindividuals would becomefitter
Figure 7: Ranking of FitnessComponents iigh,"t individualswith high penalty components' This
Number 20, Large ShaPeGDPE' sJena.iowas observedexperimentally' but is presented
of the
here only as a discussionto aid in interpretation
convergence
Plots.
ott+ G E N E T I CA L G O R I T H M S
A N DC L A S S I F I E F
SIY S T E M S

The averagevolume of the randomlygeneratedinitial It is observedthat without translocationcrossover,the


populationalso affects the convergencetrajectory. A algorithm reachesa minimum volume at about 25
heavyinitial populationwould tend to approachthe generations.Since the populationhas exhaustedits
global minimum from the top whereasa light initial diversity around generation25, the only mechanismthat
populationwould tend to approachfrom the bottom. the algorithm has left to combatthe increasingpenalty
With this said,the plots presentedin Figure 8 are meant exponentsis intelligentmutation. The increasingvolumes
to illustratethe relatively good convergence after generation 25 are therefore attributable to intelligent
characteristicswith respectto the global minimum (found mutation. Any mutationthat tendsto decreasethe shape
via an exhaustivesearchprocedure).The present penalty is accepted.The importanceof translocation
algorithm(assumingthe step-tapered cantilever crossoverand intelligent mutationis further illustratedin
configuration)achieveda minimum volume that was Figures9 and 10 where it is observedthat without
approximately2.87olarger than an exhaustivesearch translocationcrossoverand intelligentmutation,the
procedure(assuminga linearly taperingcantilever algorithm runs out of diversity at aroundgeneration20.
configuration). From Figure 8, it can be seenthat all runs Figures 9 and l0 togetherdemonstratethe ability of
(with the exceptionof a large shapepenaltyrun) achieve translocationcrossovercombinedwith intelligent
similar (favorable)results. mutation to maintain diversity during the evolutionary
process.
Figures9 and l0 illustratethe importanceof translocation
crossoverand intelligent mutationin the evolutionary Figure I I shows the aestheticallypleasingeffect of the
process.Convergencetrajectoriesare providedfor two moderategenerationallydependantshapepenaltyon the
separateruns of the algorithm. final result. The runs with large generationally
r 300 dependantshapepenaltiesalso producedsatisfactory
12 5 0 designsbut they were not as effective with respectto
1200 minimizingthe weight.This exemplifiesthe notionthat
1150 more is not always necessarybetterin terms of
o
a 1100 componentpenalties.
" t oso
o t 000
E
J 950 Generations
j soo I through 10
850
800
Generations
750 11 through 20
5 t0 15 2D 25 30 35 40 45 50
Generation

Figure9: ConvergenceTrajectories:Run #l with


Generations
ModerateShapeGDPE. 21 through 30

' |5 0 0 Generations
31 through 10
1 425

1350

o 1 2 75
o
o
Ganerations
I 200
11 through 50
1125
o
10 5 0

975

900
Evolution- MediumShape
Figure11: Cantilever
825
GDPE.
750
0 5 t0 15 20 25 30 35 40 45 50
Generation 7 CONCLUDING REMARKS
Figure 10: ConvergenceTrajectories:Run #2 with
The proposedevolutionaryalgorithm meetsall of the
ModerateShapeGDPE.
original designrequirementswith respectto the
G E N E T I CA L G O R I T H M A
SN D C L A S S I F I E R
SYSTEMS 685

theoreticaltest problem studiedhere.Given an initiai Kauf*mannPublishers,Inc.


populationof 80 randomlygeneratedl0 segment
cantilever3,where eachsegment'sx and y dimension Goldberg,D. E. (1989). GeneticAlgorithms in Search,
could take on 200 discretevalues,the algorithmwas easy Optimization,and Machine Learning. New York:
to setupso that it could consistentlycome within 37oof Addison'Wesley.
the solution found by an exhaustivesearchin 50
generations.It shouldbe emphasizedthat only 42 of the Holland, J. H. (1975).Adaptation In Natural And
80 individuals needto be evaluatedeachgenerationdue Artificial Systems,(An Introductory Analysis With
to carryoverand elitism. The resultsindicatethat the Applications To Biolog,t, Control, and Artificial
proposedevolutionaryalgorithm will scalewell and Intelligence).Cambridge,London: MIT Press.
allows a greatdeal of flexibility to deal with the
complexitiesof multi-constraintoptimization. Holland, J. H. ( 1998).Emergence: From Chaosto
Order. Reading,Mass.: Addison-Wesley.
A graphicalmethodfor interactivealgorithm tuning was
also developedwhich allows the user's intuition to be Holland, J. H. (1996).Hidden Order: How Adaption
readily incorporatedinto the selectionprocess. The rate Builds Complexity.New York: Addison-Wesley.
of convergencewas graphicallydemonstratedand easily
controlledby modifying componentsof the algorithm. Koza, J. R. (1996). GeneticProgramming: On the
Therefore,using the proposedalgorithm,the rate of Programning of Computersby Means of Natural
convergencecould be easily controlledby increasingor Selection.Cambridge:MIT Press.
decreasingthe selectivepressure.This gives the
algorithm the ability to focus on problem areaswithout Michalewicz,Z. (1992). Geneticalgorithms + data
degradingthe populationdiversity which could have structures= evolutionprograms. New York: Springer-
detrimentaleffect with a high cardinalitygenetic Verlag.
representationsuch as the heuristictree usedhere.
Mitchell, M. (1996).An Introduction to Genetic
The proposedalgorithmhas much in common with Algorithms. Cambridge:MIT Press.
simulatedannealingand was designedsuch that the final
populationcould be reheatedthru increasedintelligent Voss,M. S. and Foley,C. M. (1999),Rank Based
mutationand/or migration from multiple populations. EvolutionaryAlgorithm for StructuralOptimization.
This, combinedwith generationallydependentnon-linear Computers& Structttes (submittedfor publication).
rank basedselectionand generationallydependentpenalty
exponents allows for the implementationof iterative
improvementand multiple popuiationevolutionary
algorithmswhich could be usedto solve optimization
problemswhere the numberof constraintsis excessivefor
a singlepopulationto filter.

It should be emphasizedthat a traditionalbinary


representationis still possible,but may not be necessary
dependingon the problem and type of translocation
crossoverand mutationapplied.

Acknowledgments
The authorswould like to acknowledgethe supportof the
National ScienceFoundationruSA) - Grant Number
CMS 98 132I 6 underthe direction of Dr. Priscilla P.
Nelson. The views expressedin the paperare thoseof the
authorsand not necessarilythe sponsor.

References
Back, T. (1996). EvolutionaryAlgorithms in Theoryand
Practice.New York, New York: Oxford University Press.

Banzhaf,Nordin, Keller, and Francone(1998). Genetic


Programming - An Introduction. San Francisco:Morgan
GEGC0rgg
Troceedingoof the Geneticand
Evolul,li
onary Co mputali on Conference
A Joint Meeling of the Eighth lnternalional
on Genetic Algorithme(ICOA-?9)and
Conference
the Fourth Annual Genetic ?rogrammingConference(O?-99)

July 15-17,1'999
Orlando,Florlda

Edired by

rNolfgangbanzhaf
Jason Daida
Agoeton E. Eiben
Max H. Ganon
YagantHonavar
Mark Jakiela
Robert E. Smith

Volume1

MORGAK
] I A U F M A l I lP{ U B I . I S H E R S
S A I IF R A I I G I S G G
OA, IIFORlIIA

Das könnte Ihnen auch gefallen