《amethodtolearntheinversekinematicsofmulti-linkrobotsbyevolvingneuro-controllers.》

《《amethodtolearntheinversekinematicsofmulti-linkrobotsbyevolvingneuro-controllers.》》由会员上传分享，免费在线阅读，更多相关内容在行业资料-天天文库。

1、ARTICLEINPRESSNeurocomputing72(2009)2806–2814ContentslistsavailableatScienceDirectNeurocomputingjournalhomepage:www.elsevier.com/locate/neucomAmethodtolearntheinversekinematicsofmulti-linkrobotsbyevolving$neuro-controllersa,bcJose´AntonioMartı´nH.,JavierdeLope,MatildeSantosaDep.deSist

2、emasInforma´ticosyComputacio´n,UniversidadComplutensedeMadrid,SpainbDepartmentofAppliedIntelligentSystems,UniversidadPolite´cnicadeMadrid,SpaincDep.deArquitecturadeComputadoresyAutoma´tica,UniversidadComplutensedeMadrid,SpainarticleinfoabstractAvailableonline16April2009Ageneralmethodto

3、learntheinversekinematicofmulti-linkrobotsbymeansofneuro-controllersispresented.Wecanfindanalyticalsolutionsforthemostusedandwell-knownrobotsintheliterature.Keywords:InversekinematicsHowever,thesesolutionsarespecifictoaparticularrobotconfigurationandarenotgenerallyapplicableNeuro-evolutio

4、ntootherrobotmorphologies.TheproposedmethodisgeneralinthesensethatitisindependentoftheNeuro-controllersrobotmorphology.ThemethodisbasedontheevolutionarycomputationparadigmandworksMulti-linkrobotsobtainingincrementallybetterneuro-controllers.Furthermore,theproposedmethodsolvessomespecifi

5、cissuesinroboticneuro-controllerlearning:itavoidsanyneuralnetworklearningalgorithmwhichreliesontheclassicalsupervisedinput-targetlearningschemeandhenceitletstoobtainneuro-controllerswithoutprovidingtargets.Itcanconvergebeyondlocaloptimalsolutions,whichisoneofthemaindrawbacksofsomeneura

6、lnetworktrainingalgorithmsbasedongradientdescentwhenappliedtohighlyredundantrobotmorphologies.Furthermore,usinglearningalgorithmssuchastheneuro-evolutionofaugmentingtopologiesitisalsopossibletolearntheneuralnetworktopologywhichisacommonsourceofempiricaltestinginneuro-controllersdesign.

7、Finally,experimentalresultsareprovidedwhenapplyingthemethodtotwomulti-linkrobotlearningtasksandacomparisonbetweenstructuralandparametricevolutionarystrategiesonneuro-controllersisshown.&2009ElsevierB.V.Allrightsreserved.1.Introductionstates.Unfortunatelytheinversekinematicsproblemhas