Clustering using firefly algorithm Performance study(6)

萤火虫算法

J.Senthilnathetal./SwarmandEvolutionaryComputation1(2011)164–171

Table3

169

Table4

Table5

Table6

FAbestclassificationefficiency.

are78.7%,75.9%,89.2%,78.7%and94.2%.InthecaseofGlassandHorsedatasettheFAhasalessaverageefficiencyof70.8%and53.7%respectivelyandoverallefficiencyof61.6%and66.1%respectively.

parisonofclassificationefficiencyofnatureinspiredtechniqueusingIrisdataset

FormTable6wecanobservethat,forthestandardbenchmarkproblems—Cancer-Int,Iris,ThyroidandWinetheirisnomisclassi-ficationinanyoftheirindividualclassesi.e.allthetestdatasetareclassifiedcorrectlyandhenceCEPvalueis0.Thisdoesnotmeanthattheoverallefficiencyis100%.Toillustratethisinmorede-tail,letusconsiderIrisdatasetascomparedtoCancer-Int,Thy-roidandWine,ithaslessinputfeatures.TheFA,ABCandPSOareinthesameclassofpopulation-based,natureinspiredoptimiza-tiontechniques.Herewecomparethethreenatureinspiredtech-niquetoextractknowledgeintheformofclustercentersandtheperformanceisanalyzedusingclassificationefficiency.TheclustercentersgeneratedusingtheFA,ABCandPSOforIristrainingdataareshowninTable7.HeretheclustercentersobtainedforABCmatcheswiththepublishedliterature[35].TheparametervalueusedfortheFAisasgiveninSection5.2.1.ForABCandPSOweconsidertheparametervalueasin[9,14]respectively.

FromTable8,wecanobserveincomparisonwithothernatureinspiredtechniquestheoptimalandmeanfitnessvaluesobtainedusingtheFAisbetterthanABCandPSO.Beingacontinuousoptimizationproblem,initiallyeverypossibleclustercenterspickedbythepopulationisnotthebestoptimalpointinthesearchspace.Thereforetheselectionofnewclustercentersafterfitness