简介:OrderingbasedsearchmethodshaveadvantagesovergraphbasedsearchmethodsforstructurelearningofBayesiannetworksintermsontheefficiency.Withtheaimoffurtherincreasingtheaccuracyoforderingbasedsearchmethods,wefirstproposetoincreasethesearchspace,whichcanfacilitateescapingfromthelocaloptima.Wepresentoursearchoperatorswithmajorizations,whichareeasytoimplement.Experimentsshowthattheproposedalgorithmcanobtainsignificantlymoreaccurateresults.Withregardtotheproblemofthedecreaseonefficiencyduetotheincreaseofthesearchspace,wethenproposetoaddpathpriorsasconstraintsintotheswapprocess.Weanalyzethecoefficientwhichmayinfluencetheperformanceoftheproposedalgorithm,theexperimentsshowthattheconstraintscanenhancetheefficiencygreatly,whilehaslittleeffectontheaccuracy.Thefinalexperimentsshowthat,comparedtoothercompetitivemethods,theproposedalgorithmcanfindbettersolutionswhileholdinghighefficiencyatthesametimeonbothsyntheticandrealdatasets.
简介:Theperformanceofdeeplearning(DL)networkshasbeenincreasedbyelaboratingthenetworkstructures.However,theDLnetowrkshavemanyparameters,whichhavealotofinfluenceontheperformanceofthenetwork.Weproposeageneticalgorithm(GA)baseddeepbeliefneuralnetwork(DBNN)methodforrobotobjectrecognitionandgraspingpurpose.ThismethodoptimizestheparametersoftheDBNNmethod,suchasthenumberofhiddenunits,thenumberofepochs,andthelearningrates,whichwouldreducetheerrorrateandthenetworktrainingtimeofobjectrecognition.Afterrecognizingobjects,therobotperformsthepick-andplaceoperations.Webuildadatabaseofsixobjectsforexperimentalpurpose.Experimentalresultsdemonstratethatourmethodoutperformsontheoptimizedrobotobjectrecognitionandgraspingtasks.
简介:Weappliedthedecisiontreealgorithmtolearnassociationrulesbetweenwebpage’scategory(pornographicornormal)andthecriticalfeatures.Basedontheserules,weproposedanefficientmethodoffilteringpornographicwebpageswiththefollowingmajoradvantages:1)aweightedwindow-basedtechniquewasproposedtoestimatefortheconditionofconceptdriftforthekeywordsfoundrecentlyinpornographicwebpages;2)checkingonlycontextsofwebpageswithoutscanningpictures;3)anincrementallearningmechanismwasdesignedtoincrementallyupdatethepornographickeyworddatabase.