Genetic programming

Introduction

Geneticprogrammingisaspecialmachinelearningtechnologythatusesevolutionaryalgorithms.Itstartsfromagroupof"crowds"composedofmillionsofrandomlygeneratedcomputerprograms,andthenDeterminethesuitabilityofaprogrambasedontheabilityofaprogramtocompleteagiventask.Darwin’snaturalselection(survivalofthefittest)isusedtodeterminethewinningprogram.Computerprogramsalsosimulategendercombination,mutation,geneduplication,genedeletion,etc.Evolvefromgenerationtogenerationuntilapredeterminedstoppingconditionisreached.Thebasicideaof​​geneticprogrammingalsodrawsonthebiologicalevolutiontheoryandtheprincipleofheredityinnature,andisamethodofautomaticallyandrandomlygeneratingsearchprograms.Asanewglobaloptimizationsearchalgorithm,thisalgorithmissimple,versatile,robust,andshowsstrongabilitytosolvenonlinearcomplexproblems,soithasbeensuccessfullyappliedinmanydifferentfields,andInrecentyears,morein-depthresearchhasbeenobtained.

Applicationrangeandcharacteristics

TheapplicationrangeofGPisverywide.Intheory,anyfunctionthatobtainsavaluebasedonmultipleinputvalues,suchas:forf(x1,x2,...,xn)SuchfunctionscanbegeneratedusingGP.Whentheprogramisrelativelysimpleinlogic,itcanbedirectlywrittenbyhand,andthereisnoneedtouseGPtogenerateit,butforsomelogicallycomplexprograms,itcanbeusedtoautomaticallyevolveandgenerateaprogram.

Forexample,itisverydifficulttogeneratecontrolprogramsforagentsthathavemorecontrolresponses.Theyoftengeneratecorrespondingdecisions(actions)basedonmultipleexternalstimuli.ThistypeofprogramcanbeusedGPtogenerate.

Inthespecificimplementation,ithasthefollowingcharacteristics:

(1)WhatGPsolvesisaprogram(oranalgorithm)thatdescribestheproblem.

(2)GPisusuallyrepresentedbyatreestructure,andthedescriptionisrelativelycomplicated.

(3)Thelength(depth)ofindividualsineachgenerationofGPisgenerallydifferent,evenifthelength(depth)betweenindividualsinthesamegenerationisalsodifferent.

(4)TheresourcesconsumedbyGPareuncontrollable(uncontrollableheremeansthatitcannotbeaccuratelydescribed),itneedstoconsumealotofmemoryspace,sotheevolutionofeachgenerationisrelativelyslow.

Implementationprocess

Languageselection

Geneticprogramminghasdifferentimplementationmethodsindifferentlanguages.Theprogramgeneratedbygeneticprogrammingcanberegardedasbothprogramanddata.Intheprocessofgeneticprogramming,itisdata,whichneedstoberandomlygenerated,crossed,mutated,andevaluated;aftergeneticprogramming,itisaprogramandneedstobeexecuted.Lisplanguageisverysuitableforgeneticprogramming,becauseLisplanguagecanoperateontheprogramitselfandthenexecuteit.

Programrepresentationmethod

Programcanhavetworepresentationmethods:treestructurerepresentationandstringrepresentation.Thetreestructurerepresentationmethodhasaclearstructure.Itisrelativelysimpletoperformcrossoverandmutationoperations,butrequiresalargestoragespace.Sincegeneticprogrammingrequiresalargenumberoftreestructuresforeachgeneration,thestorageoverheadisverylarge.Spacecomplexityisatopicthatneedstobestudiedingeneticprogramming;expressingprogramswithstringscansavealotofstoragespace,butitismoredifficulttoperformgeneticoperations,suchascomparisonwhenselectingrandomintersectionsormutationpointsof"parents"Complexandneedtobeconverted.

FortheLisplanguage,itisastringitself,soitismoresuitabletouseastringtorepresentgeneticprograms;fortheClanguage,thesetwomethodscanbeselected.Sothereare4waysofcombining:Clanguage+treestructure,Clanguage+stringrepresentation,Lisp+treestructure,Lisp+stringrepresentation,ofwhichthethirdmethodisnotcommonlyused(theClanguagereferredtoheremeansinadditiontoLispMostoftheothercommonlanguages).

ItisseldomusedtoexpresstreestructureinLisplanguage,becauseitisnotnecessary.IftheClanguageisusedtorepresentthestringstructure,itissimilartotheLisplanguage,anditisrelativelysimple.YoucantreatthegeneticprogramasdataandexpresstheprogramasapseudoLisplanguage.Aftertheinheritanceisover,youcanuseitasaprogram.Ofcourse,itshouldbeexecutedasaprogramduringperformanceevaluation,whichwillbedescribedindetailbelow.

Programexecutionmethod

ForLisplanguage,programexecutionisrelativelysimple,becausethedatageneratedbygeneticsistheprogramandcanbeexecutedimmediately.FortheClanguage,itisverycomplicated,becausetheprogramsgeneratedbytheClanguagearejustdataandcannotbeexecuted.Forthisreason,aninterpretermustbedesignedtointerpretthegeneratedprogram,anddifferentinterpretersneedtobegeneratedforthetreestructureandthestringstructure.Butdifferentapplicationsneeddifferentinterpretationmethods,sotheexecutionmethodsafterinterpretationarealsodifferent.Inordertobeuniversal,anexecutioninterfaceisdesigned.Forallapplications,thesameinterpretercanbeused,whichiscalledavirtualmachine.Onlytheexecutionpartcanbechanged.Fordifferentapplications,differentexecutionfunctionsarecalled.

Evaluationfunctionselection

Fordifferentapplications,differentevaluationfunctionsneedtobeselected,soauniversalevaluationfunctioncannotbemadeintheprogramdesign.Theselectionoftheevaluationfunctioniscurrentlyasubjectthatneedstobestudied,andthereiscurrentlynouniversalandeffectivesettingmethod.Inthiscase,youcansetageneralinterface(usingthepolymorphismoftheclass),andsetdifferentevaluationfunctionsfordifferentapplications(usingasubclasstoimplementthisinterface).

Developmentprocess

ThefirstexperimentsongeneticprogrammingwerepublishedbyStephenSmith(1980)andCramm(1985).JohnKoza(1992)alsowroteafamousbooktointroducegeneticprogramming.

Computerprogramsusinggeneticprogrammingcanbewritteninmanyprogramminglanguages.Intheearly(ortraditional)implementationofGP,programinstructionsanddatavalues​​wereorganizedinatreestructure,sothoseprogramminglanguages​​thatoriginallyprovidedatreestructureweremostsuitableforGP,suchastheLisplanguageusedbyKoza.OtherformsofGPhavealsobeenpromotedandimplemented,suchasrelativelysimplelineargeneticprogrammingsuitablefortraditionalprogramminglanguages​​(suchasFortran,BASIC,andC).TherearecommercialGPsoftwarethatcombineslineargeneticprogrammingwithassemblylanguagetoobtainbetterperformance,andsomeimplementationmethodsdirectlygenerateassemblyprograms.

Theamountofcalculationrequiredforgeneticprogramming(processingalargenumberofcandidatecomputerprograms)issolargethatitcouldonlybeusedtosolvesomesimpleproblemsinthe1990s.Inrecentyears,withthedevelopmentofgeneticprogrammingtechnologyandtheexponentialincreaseinthecomputingpowerofthecentralprocessingunit,GPhasbeguntoproducealargenumberofsignificantresults.Forexample,around2004,GPachievednearly40achievementsinvariousfields:quantumcomputing,electronicdesign,gamecompetition,sorting,searchandsoon.Someofthesecomputer-generatedprograms(algorithms)areverysimilartoartificialinventionsafter2000,andeventworesultshaveresultedinnewinventionsthatcanbepatented.

Inthe1990s,itwasgenerallybelievedthatitwasverydifficulttodevelopatheoryforgeneticprogramming,andGPwasalsoatadisadvantageinvarioussearchtechnologies.After2000,thetheoryofGPhasmadesignificantdevelopment,andithasbecomepossibletoestablishanexactGPprobabilitymodelandMarkovchainmodel.Geneticprogramminghasawiderrangeofapplicationsthangeneticalgorithms(actuallyincludinggeneticalgorithms)

JuergenSchmidhuberfurtherproposedmacrogeneticprogramming,atechnologythatusesgeneticprogrammingtogenerateageneticprogrammingsystem.Somecommentsthinkthatmacro-geneticprogrammingisnotfeasibleintheory,butmoreresearchisneededtoconfirmit.