基于进化优化神经网络(EONN)的机械手运动控制(IJISA-V6-N12-2)

I.J.IntelligentSystemsandApplications,2014,12,10-16PublishedOnlineNovember2014inMECS()DOI:10.5815/ijisa.2014.12.02Copyright©2014MECSI.J.IntelligentSystemsandApplications,2014,12,10-16EvolutionaryOptimizedNeuralNetwork(EONN)BasedMotionControlofManipulatorNehaKapoorNationalInstituteofTechnology,Kurukshetra,Haryana,IndiaE-mail:[e-mail]ituteofTechnology,Kurukshetra,Haryana,IndiaE-mail:[e-mail]—Inthispaper,anEvolutionaryOptimizedNeuralNetwork(EONN)basedcontrolschemeisproposed.Thiscontrolschemeisbasedonthefactthatoptimizingvaluesofafewparametersofneuralnetworkcanenhanceitscontrolperformance.RadialBiasedNeuralNetwork(RBNN)ischosenhereandPSO,oneofthemostemergingglobaloptimizingtechniques,isusedtooptimizetheparametersofaRBNN.FromhiddentooutputlayerRBNNusesGaussianfunctionformapping.Spreadfactor(s)ofthisintelligentRBNNisthenoptimizedbyamodifiedPSOtoimproviseitsperformance.Theproposedcontrollerhasbeenverifiedbyimplementingitforpositioncontrolofaroboticmanipulator.Forcomparisonpurpose,proposedschemehasbeenverifiedwithRBNNandtheclassicalPDcontroller.MATLABenvironmenthasbeenchosenforsimulationstudycarriedout.Robustnessoftheproposedcontrollerhasbeencheckedbyapplyingittothemanipulatorforthreedifferentpaths.IndexTerms—RadialBiasNeuralNetwork(RBNN),ParticleSwarmOptimization(PSO),EvolutionaryNeuralNetwork(ENN),HybridIntelligentController,Non-LinearControlSystemsI.INTRODUCTIONRecently,twoimportantbranchesi.e.EvolutionaryAlgorithm(EA)andtheNeuralNetwork(NN)ofintelligentcontrolschemeshasbeencombined.Combinationofthesetwoimportantbranchesofcomputationalintelligence(CI)isnecessaryandinevitable[1].InspiteofagreatnonlinearlearningcapabilitiesandgoodpredictionofNeuralNetwork[2]alistofitsdisadvantagesincludeshighprocessingtime,pronetogeneralizationandtuningtheparametersofthenetworkbytrialanderrormethod;whichisatimeconsumingandfrustratingtask.TheselimitationsofNNrestrictitsuse.AutoadaptabilityinthesystemcanbeoneofthesolutionsandcanbemadeinbyusingECintelligenttechniques[3].HybridoftheseNNandEC(usedtooptimizetheparametersoftheneuralnetwork)hasfoundtobeaneffectivesolutiontothecontrolproblemsandhenceEvolutionaryNeuralNetworks(ENN)hasbeengenerated[4,5].Mostrecentlyin[6],ENNhasbeencalledasthenextgenerationNeuralNetworks.Recently,radialbasisneuralnetworks(RBNN)havebeenwidelyusedfornon-linearsystemidentification,owingtowhichRBFNNshaveonlyonehiddenlayerandhavefastconvergencespeed.Besides,theRBNNareoftenreferredtoasmodel-freeestimatorssincetheycanbeusedtoapproximatethedesiredoutputswithoutrequiringamathematicaldescriptionofhowtheoutputsfunctionallydependontheinputs.RBNNisalsoknowntobeagooduniversalapproximator.WhenutilizingRBNN,someconstantparametersmustbedeterminedfirst.AsystematicwaytodeterminetheinitialstructureofRBNNmustbeestablished.Generally,theseparametersaredeterminedaccordingtotheexperienceofthedesignerorarejustchosenrandomly.However,suchkindofimproperinitializationusuallyresultsinslowconvergencespeedandpoorperformanceoftheRBNN.In[7],RadialBiasNeuralNetworkhasbeenusedtofindtheequivalenttorqueforthemanipulatorcontrolsystem.Similarly,inthecontrolproblemdefinedinthisparticularproblem,singleinputsingleoutput(SISO)RBNNhasbeenproposed.Spreadfactor(s)istakenasaninputtotheRBNNandcontroltorqueforthesystemtobecontrolledistakenasoutput.StochasticbasedsearchalgorithmPSOhasbeenwidelyusedinrecentyearstogettheglobaloptimalsolutions[8].PSO,developedbyKennedyandElbert[9]in1995,basedonthesimulationofsimplifiedanimalsocialbehaviorsuchasfishschooling,birdflockingetc..Startingwithrandompopulationinsearchspace,itresultsintheoptimalsolution.Duringeachstepeveryparticleisacceleratedtowardsitsbestneighboringpositionaswellasinthedirectionofglobalbestposition.Calculationofnewpositionoftheswarmisgivenby(1)&(2).()()(1)(2)where,inaD-dimensionalspace⃗⃗⃗()isapresentpositionvector,⃗⃗⃗()isabestpositionvector,⃗⃗⃗()isavelocityvector,cisaconstanthavingvalue2,andaretherandomnumbergenerators.In[10,11]ithasbeenprovedthatPSOfindstheglobalbestsolution.PSOisbecomingpopularduetoitsEvolutionaryOptimizedNeuralNetwork(EONN)BasedMotionControlofManipulator11Copyright©2014MECSI.J.IntelligentSystemsandApplications,2014,12,10-16simplicityinimplementationandabilitytoconvergequicklytoareasonablygoodsolution.Roboticmanipulatorisahighlynon-linear,time-varyingandhighlycoupledsystem.Almostallkindsofcontroltechniqueshavebeendiscussedinliterature[12,13,14andreferencestherein.]foramanipulator,butbecauseofthenonavailabilityofaccuratemathematicalmodelsofmanipulatorsystems;stillthethrustforanaccurateandprecisecontrolleristhere.Moreover,manipulatorlinksaredrivenbyactuators.Actuatorsareconnectedthrougheitherdirectdriveorgeartrain.Mainlythreekindsofactuatorsareavailable:1)Pneumatic2)Hydraulictypeand3)Electricmotors.Differentactuatorshavedifferentdynamics.Electricmotorsaremorefamousbecauseoftheirlightweightandhightorque.Moreovercontrollingofelectricmotorsismucheasierthanothertwoactuators.Therefore,forhighspeedapplica