learning hierarchical control structure for multiple tasks and changing environments

《learning hierarchical control structure for multiple tasks and changing environments》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、LearningHierarchicalControlStructuresforMultipleTasksandChangingEnvironments?BruceL.DigneyRoboticsInstituteCarnegieMellonUniversityPittsburgh,PA,15232,USAPhone(412)268-7084bdigney@ri.cmu.eduofrelevanceemergelearningwillbemuchfaster.Ahier-AbstractWhiletheneedforhierarchieswithincon

2、trolsystemsarchicallearningcontrolsystemisdevelopedandtestedisapparent,itisalsocleartomanyresearchersthatagainstanon-hierarchicallearningsystem.suchhierarchiesshouldbelearned.Learningboththestructureandthecomponentbehaviorsisadicult2.Backgroundtask.Thebenetoflearningthehierarchi

3、calstruc-turesofbehaviorsisthatthedecompositionoftheAnissuecloselylinkedtorobotlearningisthearchi-controlstructureintosmallertransportablechunksal-tecturesinwhichthecontrolstrategiesareimplementedlowspreviouslylearnedknowledgetobeappliedto(Tyrrel,1992).Thetwomainideologiesare atan

4、dhi-newbutrelatedtasks.Presentedinthispaperareim-erarchical.Figure1showsthe atandthehierarchicalprovementstoNestedQ-learning(NQL)thatallowarchitecturesschematically.Flatarchitectureshavedi-morerealisticlearningofcontrolhierarchiesinrein-rectconnectionsfromsensorstoactions,througha

5、singleforcementenvironments.Alsopresentedisasimula-levelofcontrol.Whensomesituationisperceivedallbe-tionofasimplerobotperformingaseriesofrelatedtasksthatisusedtocomparebothhierarchicalandhaviorscompeteforcontrolwiththestrongestresponsenon-hierarchallearningtechniques.winning.Hiera

6、rchicalarchitectureshaveamoreindirectcouplingofperceptionstoactionsthroughahierarchical1.Introductioncontrolstructure.Whenasituationisperceived,somehighlevelbehaviorbecomesactiveandissuescommandsTheneedforhierarchicalstructureswithinlearningtooneormorelowerlevelbehavior(s).Inturn,

7、thesecontrolsystemsisclear.Withoutsomeformofhierar-lowerlevelbehaviorscontrolyetlowerlevelbehaviorschy,alearningsystemwouldbeboggeddowninthein-untilprimitiveactionsareactivatedandsomephysicalnumerabledetailsofthelowestlevelofcontrol.Manyaction(s)isperformed.Botharchitectureshavebe

8、netsresearchershaverecognizedthi