onlin knowledg dbase support vector machin abstract prior knowledg form simpl advic rule greatli speed converg learn algorithm onlin learn method predict label current point receiv correct label learn inform goal work updat hypothesi take account label feedback prior knowledg form soft polyhedr advic make increasingli accur predict subsequ exampl advic help speed bia learn gener obtain data passiv aggress approach updat hypothesi hybrid loss take account margin hypothesi advic current point encourag comput result loss bound provid transfer learn advic take abstract goal transfer learn speed learn task transfer knowledg relat sourc task describ transfer method reinforc learner analyz experi sourc task learn rule advic target task rule learn induct logic program describ condit action success sourc task advic take algorithm target task reinforc learner benefit rule imperfect human provid map describ align sourc target task includ advic differ task robocup simul soccer domain demonstr transfer method speed reinforc learn substanti rule extract transfer learn typic rule extract purpos human interpret applic rule extract practic applic transfer learn knowledg learn task aid learn relat task extract rule explain learn solut relat macro transfer reinforc learn abstract describ applic induct logic program transfer learn transfer learn knowledg learn sourc task improv learn relat target task task work reinforc learn domain approach transfer relat macro finit state machin transit condit node action repres order logic claus induct logic program learn macro character success behavior sourc task macro decis make earli learn stage target task experi robocup simul soccer domain show relat macro transfer demonstr rmt sourc task provid substanti head start target task build relat world model reinforc learn abstract reinforc learn domain highli relat tradit tempor differ method appli domain limit capac exploit relat natur domain algorithm ambil construct relat world model form relat markov decis process mdp ambil work backward collect high reward state util induct logic program learn preimag logic definit region state space lead high reward state action learn preimag chain form mdp abstractli repres domain ambil estim reward transit probabl mdp past experi mdp small ambil valueiter quickli estim valu action induc state determin polici ambil emploi complex background knowledg support relat represent empir evalu synthet domain task robocup soccer domain show signific perform gain compar standard learn simpl effect method incorpor advic kernel method propos simpl mechan incorpor advic prior knowledg form simpl rule support vector method classif regress approach base introduc inequ constraint datapoint match advic constrain datapoint standard exampl train set unlabel data semi supervis advic take approach approach simpler implement effici solv knowledg base support vector classif method fung mangasarian shavlik knowledg base support vector regress method mangasarian shavlik wild perform approxim complex approach experi approach synthet task reinforcementlearn problem robocup soccer simul show advic take method significantli outperform method advic perform similarli prior advic take support vector machin end dto dend supervis target dword sens disambigu system present extens supervis target word sens disambigu system leverag gate kernel method word sens disambigu acronym expans skill acquisit transfer learn advic take abstract describ reinforc learn system transfer skill previous learn sourc task relat target task system induct logic program analyz experi sourc task transfer rule action target task learner accept rule advic take algorithm learner benefit guidanc imperfect system accept human provid map specifi similar sourc target task includ advic differ task robocup simul soccer domain demonstr system speed reinforc learn substanti give advic prefer action reinforc learner knowledg dbase kernel regress present formul provid advic reinforc learner emploi supportvector regress function approxim method extend recent advic give techniqu call knowledg base kernel regress kbkr accept advic singl action reinforc learner kbkr user set state action greater linear express current state techniqu call prefer kbkr pref kbkr user provid advic natur manner recommend action prefer set state prefer essenti mean user give advic polici valu natur human present advic present motiv prefer advic proof correct extens kbkr addit show empir result method make effect advic reinforc learn task base robocup simul call breakawai work demonstr signific potenti advic give techniqu address complex reinforc learn problem demonstr support vector regress reinforc learn advic transfer knowledg acquir reinforc learn task abstract present method transfer knowledg learn task relat task problem solver emploi reinforc learn acquir model task transform learn model advic task human teacher map task task guid knowledg transfer advic incorpor problem solver knowledg base support vector regress method previous develop advic take approach problem solver refin discard transfer knowledg base subsequ experi empir demonstr effect approach game robocup soccer simul keepawai breakawai result demonstr problem solver learn plai breakawai advic extract keepawai outperform problem solver learn benefit advic compar studi support vector machin appli supervis word sens disambigu problem medic domain abstract appli supervis learn approach word sens disambigu medic domain object evalu support vector machin svm comparison supervis learn algorithm includ na ve bay classifi decis tree decis list boost approach base result introduc refin approach made unigram bigram featur select frequenc cut valu window size statist signific test log likelihood measur bigram result show svm model accur case compar na ve bay decis tree decis list boost method knowledg transfer advic take exploit unlabel data ensembl method adapt semi supervis ensembl method assembl propos construct classif ensembl base label unlabel data assembl altern assign pseudo class unlabel data exist ensembl construct base classifi label pseudolabel data mathemat intuit algorithm correspond maxim classif margin hypothesi space measur label unlabel data unlik altern approach assembl requir semi supervis learn method base classifi assembl conjunct cost sensit classif algorithm class multi class problem assembl decis tree won nip unlabel data competit addit strong result benchmark dataset decis tree neural network support propos method popular ensembl method empir studi ensembl consist set individu train classifi neural network decis tree predict combin classifi instanc previou research shown ensembl accur singl classifi ensembl bag breiman boost freund amp schapir schapir popular method produc ensembl paper evalu method data set neural network decis tree classif algorithm result number conclus bag accur singl classifi accur boost hand boost creat ensembl accur singl classifi neural network analysi perform boost method depend characterist data set examin fact result show boost ensembl overfit noisi data set decreas perform final consist previou studi work suggest gain ensembl perform classifi combin larg gain classifi boost decis tree boost classifi region paper present algorithm boost perform ensembl classifi boost seri classifi predict class data member seri concentr train data incorrectli predict earlier member make predict pattern classifi predict class pattern predict combin standard boost predict combin weight predict term relat accuraci classifi train data approach ignor fact classifi focu small subset pattern good classifi similar pattern regionboost problem address weight classifi predict factor measur classifi perform similar pattern paper examin method determin classifi perform simil empir evalu bag boost ensembl consist set independ train classi er neural network decis tree predict combin classifi instanc previou research shown ensembl accur singl classier ensembl bag breiman boost freund amp schapir popular method produc ensembl paper evalu method neural network decis tree classi cation algorithm result showtwo import fact rst bag produc classi er individu compon classi er impervi tting gener baselin neural network ensembl method boost apow techniqu produc ensembl bag suscept nois quickli data set featur engin classifi select case studi venusian volcano detect machin learn graduat toi problem real world applic user find real world problem requir perform aspect problem solv address machin learn literatur specif user find task select set featur defin problem obtain set exampl problem import success machin learn applic select develop specif classif method paper present case studi machin learn appli difficult real world problem detect volcano sar synthet apertur radar imag venu magellan dataset work demonstr process featur select sampl collect critic product good classifi show domain depend knowledg serv enhanc result ensembl sequenc classifi ensembl classifi creat combin predict multipl compon classifi present method combin classifi ensembl base simpl estim classifi compet classifi group order list classifi threshold classifi classifi list consult classifi confid predict classifi threshold classifi predict classifi threshold consult classifi predict confid threshold class predict averag compon classifi predict kei method select confid threshold classifi implement method system call sequel appli task recogn volcano sar imag venu domain sequel outperform individu classifi simpl approach ensembl construct averag predict classifi combin predict multipl classifi competit learn initi neural network primari goal induct learn gener induc function accur produc correct output futur input hansen salamon show assumpt combin predict separ train neural network improv gener kei assumpt individu network independ error produc standard perform backpropag assumpt violat standard procedur initi network weight region weight space origin mean backpropag gradient descent search reach small subset local minima paper present approach initi neural network competit learn intellig creat network origin locat origin weight space potenti increas set reachabl local minima creat advic dtake reinforc learner learn reinforc promis approach creat intellig agent reinforc learn requir larg number train episod present evalu design address shortcom allow connectionist learner accept advic time natur manner extern observ approach advic giver watch learner occasion make suggest express instruct simpl imper program languag base techniqu knowledg base neural network insert program directli agent util function subsequ reinforc learn integr refin advic present empir evid investig aspect approach show good advic learner achiev statist signific gain expect reward experi show advic improv expect reward combin predict multipl classifi competit learn initi neural network primari goal induct learn gener induc function accur produc correct output futur input hansen salamon show assumpt combin predict separ train neural network improv gener kei assumpt individu network independ error produc standard perform backpropag assumpt violat standard procedur initi network weight region weight space origin mean backpropag gradient descent search reach small subset local minima paper present approach initi neural network competit learn intellig creat network origin locat origin weight space potenti increas set reachabl local minima incorpor advic agent learn reinforc learn reinforc promis approach creat intellig agent reinforc learn requir larg number train episod present approach address shortcom allow connectionist learner accept advic time natur manner extern observ approach advic giver watch learner occasion make suggest express instruct simpl program languag base techniqu knowledg base neural network program insert directli agent util function subsequ reinforc learn integr refin advic present empir evid show approach lead statist signific gain expect reward importantli advic improv expect reward stage train introduct success increasingli popular method cr knowledg dbase neural network improv algorithm refin chou dfasman algorithm protein fold describ method machin learn refin algorithm repres gener finit state automata knowledg automaton translat artifici neural network refin backpropag set exampl techniqu translat automaton network extend kbann system translat set proposit rule neural network extend system fskbann refin larg class algorithm repres state base process test fskbann refin chou fasman algorithm method predict globular protein fold empir evid show refin algorithm fskbann produc statist significantli accur origin chou fasman algorithm neural network train standard approach introduct machin learn increasingli appli complex real world problem research knowledg dbase neural network improv algorithm refin chou dfasman algorithm protein fold describ method machin learn refin algorithm repres gener finit state automata knowledg automaton translat artifici neural network refin backpropag set exampl techniqu translat automaton network extend kbann system translat set proposit rule neural network extend system fskbann refin larg class algorithm repres state base process test fskbann refin chou fasman algorithm method predict globular protein fold empir evid show refin algorithm fskbann produc statist significantli accur origin chou fasman algorithm neural network train standard approach introduct machin learn increasingli appli complex real world problem research refin domain theori express finit dstate automata abstract kbann system neural network torefin domain theori domain knowledg kbann express recurs proposit rule extend kbann domain theori express asfinit state automata appli finit state kbann task predict proteinsfold produc small statist signific gain accuraci astandard neural network approach learn algorithm biologicalliteratur method show promis solv real world problemsthat term statedepend decis introduct research artifici neural network ann untilrec larg preexist knowledg enrich vocabulari gener explan structur 