欢迎来到天天文库
浏览记录
ID:38125317
大小:163.40 KB
页数:6页
时间:2019-05-27
《Incremental rule-based control and learning》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、Incrementalrule-basedcontrolandlearningD.LuzeauxEECSDepartmentUniversityofCaliforniaBerkeley,CA94720E-mail:luzeaux@robotics.berkeley.edu,luzeaux@etca.frAbstractThesecontrollawsarefurthermorerathereasytolearn:thisisCandide'sjobandwegiveinthepa-Autonomoussystemsarenowadaysachalle
2、ngeperthelearningmethodologyandshowhowCandidetomanyclassical"techniques:theyareeitherill-learnedtodriveacar.modeledorevolveinchangingenvironmentswhichperturbatethemdeeply.Itisnotpossibletoapplythewell-knownmathematicaltoolsandnewmethodshave2Incrementalrule-basedcontrollawstobe
3、foundtocontrolthem.Thesenewmethodsareusuallyknownasintelligentcontrolandincludeneu-Thesecontrollaws,whosemathematicalformula-ralnetworks,fuzzycontrol,qualitativecontrol,path-tionfollows,havebeenstudiedexperimentallyduringplanningmethodsandrule-basedreasoning.thelastyears[1,4,2]
4、;howevertheoreticalresultsareveryrecent[6,7].Inthispaperwerstintroduceincrementalrule-basedcontrollawsandthenwediscussalearningpro-2.1Theoreticalresultsgram,calledCandide,whichallowsalargeclassofsystemstosynthesizesuchcontrollawsinordertoper-formagiventask.Theincrementalrule-b
5、asedcontrollawshavethefollowingform:Ifoutputinthisrangethenincrease,decreaseorleavetheinput1IntroductionTheinputuattimekisrelatedtotheinputuatkk?1timek?1bytheequation:Whychooseincrementalrule-basedcontrollaws?8Theanswermayseemtorelymostlyonintuition0100u=u+u+>kk?1kkk>>butrec
6、enttheoreticalresults[10,6]showtheideawas1112>u=u+u>kk?1kk>n?1n?1n?1n>>u=u+uknowledgeinmostcasesinadeclarativeform.Iftheykkk?1k>:nu=havetoactonthesystem,theydonottrytosolvekacomplexe
7、quationandthenfeedthesolutiontothei0with:8i2f0;;n?1g;8k2N;2f?1;0;1g;2system:theyusuallyincreasetheirpreviousactionorkk0f?m;;mg,and;aretwopositiverealnum-decreaseitbysomeamountortheykeepitthesame.i0bers.ThecoecientsandwillbecalledsignskkTheincrementalrule-basedcont
8、rollawsarethefor-andthesetoftheirvaluesattimekisthesig
此文档下载收益归作者所有