Abstract — An attempt is made to apply Ant Colony Optimization (ACO) heuristics to an E-learning problem: the pedagogic material of an online teaching website for high school students is modelled as a navigation graph where nodes are exercises or lessons

5
4
3
Fig.1.Theelementarytestcaseisasituationthatcouldbeseenasinappropriate:Arc1-3isfavouredbythepedagogicalteam,whereasthesuccessratefornode4ismuchhigherwhentheantscomefromnode2thanwhentheycomefromnode3.
2
1Legend
fitness of arc 1–3fitness of arc 1–2solvingexercice1,thestudentcaneithergotoexercise2orexercise3.Exercise3isencouragedbythepedagogicalteamasthearcleadingtoitisassignedaweightof5versus1forthearcleadingtoexercice2,whichhoweveryieldsa vetimesgreaterprobabilityforstudentstobesenttoexercise3afterexercise1isvalidated.
Theproblemisthatthesuccessrateofexercice4ismuchhigherwhenthestudentcomesfromexercise2thanwhenhe/shecomesfromexercise3.Whatisexpectedfromthesysteminsuchacaseistodetectthesituationandtoinversethetwoprobabilitiessothatstudentsareencouragedtofollowtherightpath.Thisshouldbeachievednaturally,i.e.withoutanyhumanintervention,thankstothereleaseofvirtualpheromonesalongthearcs.
Thearcleadingtoexercise2aregoingtoholdalargeamountofsuccesspheromonesandalowamountoffailurepheromones.Thearcleadingtoexercise3,onthecontraryisgoingtobeintheoppositesituationandthisdoubledis-crepancyisgoingtobere ectedinthearcs’ tnesses,therebymodifyingtheirprobabilitiestobefollowed.Progressively,thearcleadingtoexercise2isgoingtotakeoverthearcleadingtoexercise3andareasonablesituationshouldbepromptlyreestablished.
Thenextsectiondescribesexperimentsthatwereconductedtomakesurethatthesystemwasabletobehaveasexpected,andtodeterminethecorrespondingappropriaterangeofparameters.
VII.FIRSTRESULTS
Thissectionreportsexamplesofnumericalresultsobtainedaftercalibrationexperimentsconductedonthereducedgraph.A.GeneralCalibrationforStability
The rstlessondrawnfrompreliminaryexperimentsisthatwithoutcarefulcalibration,thesystemcanbehaveinavarietyofundesirablewaysthatincludes tnessvaluesquicklydivingtozero(toostrongevaporation),randomlygrowingtoin nity(toostrongαs)oroscillatingaroundtheiroriginalvalues.Thelattercaseisillustratedbytheexampleshownherewherethe
Fig.2.Fitnessesofarcs(1,3)and(1,2)versusthenumberofiterations.τ=0.99Oscillatingbehaviour,T doesnotexist.(αs=αf=α=0.1,ω1=ω2=1,ω3= 3)
evaporationrateτiscalibrated.Bothcurves(see gures2and3)showtherespective tnessesofarcs(1,3)and(1,2)plottedversusthenumberofiterations,i.e.thenumberoftimesanantgoesfromonenodetoanother(whenanantreachestheendofthegraph,anewoneisplacedatthestartingnode).Inthe rstcase( gure2),withτ=0.99,evaporationistooquickandthearc tnessesdon’thavetimetogetinvertedastheyshould.
Theirvalueskeeposcillatingpulledupbypheromonesandpulledbackdownbyevaporation.Withahigherτ=0.999,thesecondcurve( gure3)showstheappropriateexpectedbehaviour: tnessesarequicklyinverted(ataniterationwewillnameT andremaininverted,eitherstabilising(f(1,3))orgrowingatareasonablerate(f(1,2)).
Inthisparticulartestcase,thesystemisconsideredtobestablewhenitbehavesthisway.ItmeansthatT ,theinversioniteration,existsandthatthe tnessvaluesf(1,3)andf(1,2)remainproperlyinvertedbeyondT .B.Re nedParameterTuning
T appearstobeakeyparameterofthesystem,describingthevelocitywithwhichappropriatechangesaremadetothearc tnessdistribution.ThefollowingsetofexperimentsaimsatunderstandinghowT isin uencedbythesystem’svariousparameters.Thisshouldallowforcompetenttuningofthoseparameters.
Curvesshownin gures4and5plotthevalueofT averagedoverseveralrunsof1001iterationsversusarangeofvaluesforthestudiedparameter.Twoparameters,identi edaskeyonesareinvestigated:α,thequantumofreleasedpheromone(fortheseexperiments,α=αs=αf)andω3
热门文档