inform lift messag dpass lift infer handl set indistinguish object critic effect applic probabilist relat model realist real world task recent lift belief propag lbp propos effici approxim solut infer problem run modifi bp lift network node group roughli speak ident comput tree tree structur unrol underli graph root node situat pure syntact criterion pessimist messag error decai path intuit long chain graph weak edg potenti distant node send receiv ident messag comput tree overcom propos ilbp easi implement inform lbp approach interleav lift modifi bp iter turn effici monitor true bp messag receiv iter group node experi show ilbp yield significantli faster lift network degrad perform show ilbp faster bp solv problem distribut data larg network import real world applic bp faster uninform lbp symbol dynam program dorder pomdp partial observ markov decis process pomdp provid power model sequenti decis make problem partial observ state approxim optim dynam program solut work recent year focus improv effici dynam program algorithm exploit symmetri factor relat represent work show exploit full express power order quantif achiev state action observ abstract dynam program solut relation pomdp advantag approach abil maintain compact function represent abstract space potenti optim action automat deriv compact condit polici tree minim partit relat observ space distinct impact polici valu lift relat pomdp solut optim accommod action potenti infinit relat space observ outcom simplex volum maxim descript web dscale matrix factor matrix factor method common techniqu detect latent compon data popular exampl includ singular decomposit nonneg matrix factor method suffer high comput complex scale massiv data paper present linear time algorithm factor gigant matric iter yield latent compon constrain matrix factor latent compon form simplex enclos remain data algorithm maxim volum simplex reduc displac data space span latent compon lower frobeniu norm common criterion matrix factor qualiti algorithm effici ground distanc geometri easili applic matric billion entri addit result factor intuit interpret data data point express convex combin extrem descript instanc collect data extens experiment valid web scale data includ million imag million twitter tweet demonstr superior perform compar relat factor cluster techniqu explor relat world abstract kei problem model base reinforc learn balanc explor exploit learn act larg relat domain vari number object relat provid solut explor larg relat markov decis process develop relat extens concept explicit explor exploit algorithm kei insight inher gener learnt knowledg relat represent profound implic explor strategi proposit set consid situat worth explor relat set instanc context exploit promis experiment evalu show effect benefit relat explor proposit benchmark approach noisi simul robot manipul problem exploit causal independ markov logic network combin undirect direct model abstract method propos compil causal independ markov logic network mln mln view compactli repres factor joint probabl product set factor guid logic formula present notion causal independ enabl factor factor combin smaller factor obtain finer grain factor joint probabl causal independ let factor term weight direct claus oper sum max contribut variabl involv factor combin undirect direct knowledg experiment evalu show make finer grain factor provid causal independ improv qualiti paramet learn mln exploit causal independ markov logic network combin undirect direct model abstract method propos compil causal independ markov logic network mln mln view compactli repres factor joint probabl product set factor guid logic formula present notion causal independ enabl factor factor combin smaller factor obtain finer grain factor joint probabl causal independ let factor term weight direct claus oper sum max contribut variabl involv factor combin undirect direct knowledg experiment evalu show make finer grain factor provid causal independ improv qualiti paramet learn mln topic model condit relat abstract latent dirichlet alloc fulli gener statist languag model proven success captur content topic corpu document recent shown relat document hyper link citat share inform document turn improv topic gener fulli gener situat interest predict relat document paper present dirichlet multinomi nonparametr regress topic model includ gaussian process prior joint document topic distribut function document relat network scientif abstract wikipedia document show approach meet exce perform baselin topic model fast activ explor link dbase prefer learn gaussian process abstract prefer learn algorithm observ pairwis rel judgment prefer item train data learn order item import learn problem applic absolut feedback difficult elicit pairwis judgment readili implicit feedback shown activ learn effect reduc number train pair need success exist algorithm gener item queri web search learn separ relev score document queri pair scratch overcom ineffici propos link base activ prefer learn method base gaussian process gp incorpor depend inform featur vector represent relat specif meet requir comput effici activ explor introduc increment updat method scale gener model propos algorithm evalu dataset inform retriev show learn substanti faster algorithm model depend kernel map match dtaught decis theoret plan order decis diagram dgrid depend maxim view imag brows learn prefer hidden common relat abstract gaussian process successfulli learn prefer entiti provid nonparametr bayesian approach model select probabilist infer entiti encount real world applic complex relat paper present prefer model incorpor inform relat entiti specif propos probabilist relat kernel model prefer learn base silva al mix graph gaussian process prior distribut enhanc relat graph kernel propos captur correl prefer empir analysi letor dataset demonstr relat inform improv perform prefer learn stack gaussian process learn convex dneg matrix factor wild abstract neg matrix factor nmf recent receiv lot attent data mine inform retriev comput vision factor neg input matrix neg matrix factor wh describ cluster dataset analyz genotyp social network imag benefici ensur meaning cluster centroid restrict convex combin data point run convex nmf wild million data point trigger simpl observ data point convex combin vertic data convex hull propos restrict vertic convex hull benefit convex hull nmf approach twofold expect size convex hull random gaussian point plane log candid set typic grow slower data set distanc preserv low dimension embed comput candid vertic effici extens experiment evalu show convex hull nmf compar favor convex nmf larg data set term speed reconstruct qualiti show method easili appli larg scale real world data set case consist million imag million vote world warcraft guild keyword data mine matrix decomposit data handl neg matrix factor archetyp analysi social network analysi kernel condit quantil estim reduct revisit abstract quantil regress refer process estim quantil condit distribut import applic econometr data mine domain paper show estim condit quantil function bay risk minim framework gaussian process prior result parametr probabilist model easi implement cross quantil function enforc directli combin tool extens standard gaussian process principl hyperparamet estim sparsif quantil regress input depend nois rate exist approach enjoi desir properti experi benchmark dataset show method competit state art approach keyword regress quantil regress gaussian process multi drelat learn gaussian process due flexibl nonparametr natur gaussian process model effect solv hard machin learn problem exist gaussian process model focu model singl relat present gener gp model name multi relat gaussian process model deal arbitrari number relat domain interest propos model analyz context bipartit direct undirect univari relat experiment result real world dataset show exploit correl entiti type relat improv predict perform gener order decis diagram order markov decis process order decis diagram fodd recent introduc compact knowledg represent express function relat structur fodd repres numer function constrain boolean rang existenti quantif previou work develop set oper fodd show solv relat markov decis process rmdp dynam program algorithm demonstr success solv stochast plan problem intern plan competit system fodd planner crucial ingredi scheme set oper remov redund decis diagram keep compact paper make contribut introduc gener fodd gfodd combin algorithm gener fodd arbitrari quantif show gfodd principl solv rmdp arbitrari quantif develop promis case arbitrari number existenti quantifi arbitrari number univers quantifi develop approach reduc fodd gfodd model check yield reduct complet fodd sound reduct procedur gfodd ilp blind eleph euclidean embed dproven queri abstract relat data complex complex make basic step ilp difficult understand data result user easili understand draw incomplet conclus situat parabl blind men eleph appear cultur tale blind work independ piec inform draw conclus natur beast contrast visual represent make easi shift perspect explor analyz data paper describ method embed interpret queri singl common euclidean space base proven statist demonstr method real world dataset show ilp result captur glanc ilp blind eleph euclidean embed dproven queri abstract relat data complex complex make basic step ilp difficult understand data result user easili understand draw incomplet conclus situat parabl blind men eleph appear cultur tale blind work independ piec inform draw conclus natur beast contrast visual represent make easi shift perspect explor analyz data paper describ method embed interpret queri singl common euclidean space base proven statist demonstr method real world dataset show ilp result captur glanc probabilist abstract paper present bayesian method construct probabilist network databas focu construct bayesian belief network potenti applic includ comput assist hypothesi test autom scientif discoveri autom construct probabilist expert system extend basic method handl miss data hidden latent variabl show perform probabilist infer averag infer multipl belief network result present preliminari evalu algorithm construct belief network databas case final relat method paper previou work discuss open problem probabilist induct logic program theori applic lift probabilist infer count formula lift infer algorithm exploit repeat structur probabilist model answer queri effici previou work de salvo braz al order variabl elimin fove focus share potenti interchang random variabl paper exploit interchang individu potenti introduc count formula random variabl set present lift infer algorithm fove handl count formula input creat count formula intermedi potenti fove succinctli term oper heurist appli count formula captur depend larg number variabl compactli fove achiev asymptot speed improv compar fove paramet learn probabilist databas squar approach abstract introduc problem learn paramet probabilist databas problog observ success probabl set queri comput probabl attach fact low approxim error train exampl unseen exampl assum gaussian error term observ success probabl natur lead squar optim problem approach call leproblog learn queri proof simultan make flexibl faster train domain proof experi real world data show us effect squar calibr probabilist databas nonstationari gaussian process regress point estim local smooth abstract gaussian process nonstationari covari function power tool bayesian regress input depend smooth common approach model local smooth latent process integr markov chain mont carlo approach paper demonstr approxim estim local smooth yield good result emploi effici gradient base optim techniqu jointli learn paramet latent observ process extens experi synthet real world data includ challeng problem robot show relev feasibl approach boost relat sequenc align task align sequenc aris applic classic dynam program approach requir explicit state enumer reward model impract number state grow quickli number domain object relat object relat sequenc align aim exploit symbol structur avoid full enumer expens complex reward model select problem virtual infinit abstract level explor paper appli gradientbas boost leverag problem specif show reduc learn problem seri relat regress problem main benefit interact state variabl introduc need potenti infinit search space explicitli consid experiment result show boost approach significantli improv establish result challeng applic dparametr polici gradient unifi treatment proposit relat domain polici gradient approach power instrument learn interact environ exist approach focus proposit continu domain extens featur engin difficult imposs appli structur domain vari number object relat paper describ parametr polici gradient approach call nppg overcom limit kei idea appli friedmann gradient boost polici repres weight sum regress model grown stage wise optim emploi shelf regress learner nppg deal proposit continu relat domain unifi experiment result show improv establish result logic hierarch hidden markov model model user activ abstract hidden markov model hmm successfulli applic speech recognit activ recognit bioinformat previou attempt hierarch hmm abstract hmm elegantli extend hmm multipl level tempor abstract repres user activ similarli previou work logic hmm extend hmm domain relat structur work develop represent natur combin power relat hierarch model form logic hierarch hidden markov model lohihmm lohihmm inherit compact represent logic hmm tractabl infer hierarch hmm outlin infer algorithm base ground lohihmm proposit hmm base particl filter adapt set present result experi model simul domain srl tear ilp perspect learn predict terrain model leg robot locomot abstract leg robot requir accur model environ order plan execut path present probabilist techniqu base gaussian process terrain model learn updat effici spars approxim techniqu major benefit terrain model abil predict elev unseen locat reliabl altern approach yield estim uncertainti predict nonstationari gaussian process model adapt covari situat hand allow accur infer terrain height point observ directli show convent motion planner learn terrain model plan path goal locat terrain specif cost model accept reject candid foothold experi real quadrup robot equip laser rang finder demonstr us approach discuss benefit compar simpler terrain model elev grid engag game probabilist induct logic program abstract paper present probabilist method deal multiclass classif stochast logic program slp probabilist induct logic program pilp framework integr probabl logic represent learn multi class predict attempt classifi observ datum proper classif test multipl predict appli slp paramet estim algorithm previou studi protein fold predict area multi class classif work logic program learn induct logic program ilp larg number multipl predict detect basi experi demonstr pilp approach slp advantag solv multi class predict problem learn probabl addit show slp outperform ilp major class predictor predict accuraci result interpret basic principl learn bayesian logic program bayesian logic program tightli integr definit logic program bayesian network order paper present result combin induct logic program bayesian network learn qualit quantit compon bayesian logic program data precis show qualit compon learn combin induct logic program set learn interpret score base techniqu learn bayesian network estim quantit compon reduc problem dynam bayesian network relat sequenc learn abstract wetackletheproblemofmulti classrel sequencelearn relev pattern discov set label sequenc deal problem firstli relat sequenc map featur vector result featur construct method efficaci sequenc learn algorithm strongli depend featur repres sequenc step find optim subset construct featur lead high classif accuraci featur select task solv adopt wrapper approach stochast local search algorithm embed na ve bay classifi perform propos method appli real world dataset show improv compar establish method hidden markov model fisher kernel condit random field relat sequenc social network mine nonparametr relat model abstract statist relat learn srl effect techniqu analyz social network data rich collect object complex network infinit hidden relat model ihrm introduc nonparametr mixtur model relat learn success relat applic paper explor model analysi complex social network ihrm commun detect link predict product recommend ihrm base social network model edg random variabl probabilist depend random variabl model base relat structur hidden variabl object transport inform local probabilist depend obtain model predict entiti attribut predict relationship entiti perform interpret cluster analysi demonstr perform ihrm social network applic perform commun analysi sampson monasteri data perform link analysi bernard amp killworth data final appli ihrm movielen data predict user prefer movi analysi user cluster movi cluster prefac part seri studi support california governor advisori committe educ excel prefac list figur iv compress probabilist prolog program mine learn graph link base data studi formal mean unord tree dataset form link base data natur notion support base closur immedi defin abstract inform subset data requir formal notion intersect deeper understand notion closur cient algorithm comput intersect unord tree provid answer question heteroscedast gaussian process regress paper present gaussian process gp approach regress inputdepend nois rate follow goldberg al approach model nois varianc gp addit gp govern nois free output contrast goldberg al markov chain mont carlo method approxim posterior nois varianc nois approach result model easi implement directli combin exist extens standard gp spars approxim extens experi synthet real world data includ challeng percept problem robot show effect heteroscedast gp regress abstract collect probabilist gaussian beam process nonparametr bayesian measur model rang finder abstract probabilist mobil robot develop measur model plai crucial role directli influenc effici robust robot perform great varieti task includ local track map build paper present probabilist measur model rang finder call gaussian beam process treat measur model task nonparametr bayesian regress problem solv gaussian process major benefit approach abil gener entir rang scan directli learn distribut rang measur region robot configur space record simul rang scan approxim approach state estim particl filter histogram filter lead approxim true likelihood function experi real world synthet data show gaussian beam process combin advantag popular measur model fisher kernel relat data abstract combin statist relat learn receiv lot attent major statist relat learn approach focu densiti estim classif perform gener model lower discrimin classifi approach improv perform gener model combin discrimin algorithm fisher kernel develop combin kernel method shown promis result combin support vector machin logic hidden markov model bayesian network fisher kernel consid relat data data consist collect object relat object paper develop fisher kernel relat data empir show significantli improv result achiev fisher kernel tildecrf condit random field logic sequenc abstract condit random field crf provid power instrument label sequenc crf consid label sequenc flat alphabet paper describ tildecrf method train crf logic sequenc sequenc alphabet logic atom tildecrf kei idea relat regress tree dietterich al gradient tree boost approach crf potenti function repres weight sum relat regress tree experi show signific improv establish result achiev hidden markov model fisher kernel logic sequenc robust scan point classif associ markov network abstract paper present effici techniqu learn associ markov network amn segment scan data techniqu extens work recent present anguelov al amn appli learn max margin optim paper show adapt reduc train data train process perform effici achiev good classif result reduct obtain util kd tree prune appropri algorithm requir addit paramet yield abstract train data experi real data collect mobil outdoor robot demonstr approach yield accur segment relat sequenc align logo revis probabilist prolog program abstract problog probabilist prolog languag introduc algorithm develop solv approxim problog queri defin studi problem revis problog theori exampl problog probabilist prolog problog program consist prolog set definit claus problog claus ci label probabl pi true bibliograph data analysi similar structur item improv inform retriev result collect paper pairwis similar similar base kei word analysi item relat similar similar item relat uncertainti data infer elegantli repres attach probabl learn relat navig polici navig fundament task mobil robot major path plan approach design solv problem scratch current goal configur robot approach yield highli effici plan comput polici typic transfer similar task propos learn relat decis tree abstract navig strategi path relat abstract interest import properti mobil robot gener navig plan specif exampl provid user explor navig polici learn environ transfer unknown environ experi real robot real environ simul run demonstr us approach logic hidden markov model logic hidden markov model lohmm upgrad tradit hidden markov model deal sequenc structur symbol form logic atom flat charact note formal introduc lohmm present solut central infer problem lohmm evalu hidden state sequenc paramet estim result represent algorithm experiment evalu problem domain bioinformat induct logic program approach statist relat learn logic hidden markov model logic hidden markov model lohmm upgrad tradit hidden markov model deal sequenc structur symbol form logic atom flat charact note formal introduc lohmm present solut central infer problem lohmm evalu hidden state sequenc paramet estim result represent algorithm experiment evalu problem domain bioinformat learn stochast logic program proof dbank stochast logic program combin idea probabilist grammar express power definit claus logic consid extens probabilist context free grammar motiv analog learn tree bank grammar studi learn stochast logic program proof tree proof tree exampl impos strong logic constraint structur target stochast logic program constraint integr gener gener lgg oper emploi travers search space implement emploi greedi search guid maximum likelihood principl failur adjust maxim report number simpl experi show promis approach nfoil integr na efv bay foil em select probabilist model logic sequenc real world sequenc protein secondari structur shell log exhibit rich intern structur tradit probabilist model sequenc sequenc flat symbol logic hidden markov model propos solut deal logic sequenc sequenc alphabet logic atom expens complex model select problem abstract level explor paper propos method select logic hidden markov model data call sagem sagem combin gener expect maxim optim paramet structur search model select induct logic program refin oper provid converg experiment result show sagem effect probabilist induct logic program abstract paper present probabilist method deal multiclass classif stochast logic program slp probabilist induct logic program pilp framework integr probabl logic represent learn multi class predict attempt classifi observ datum proper classif test multipl predict appli slp paramet estim algorithm previou studi protein fold predict area multi class classif work logic program learn induct logic program ilp larg number multipl predict detect basi experi demonstr pilp approach slp advantag solv multi class predict problem learn probabl addit show slp outperform ilp major class predictor predict accuraci result interpret fisher kernel logic sequenc abstract approach improv accuraci classif base gener model combin success discrimin algorithm fisher kernel develop combin gener model popular class learn algorithm kernel method empir combin hidden markov model support vector machin shown promis result fisher kernel consid sequenc flat alphabet due lack method comput gradient gener model structur sequenc paper show comput gradient logic hidden markov model model logic sequenc sequenc alphabet logic atom experi show consider improv result achiev fisher kernel logic sequenc bellman relat motiv interest relat reinforc learn introduc relat bellman updat oper call rebel emploi constraint logic program languag compactli repres markov decis process relat domain logic markov decis program converg logic td lambda recent develop area relat reinforc learn rrl result number algorithm theori explain rrl work lack paper provid initi result theori rrl realiz introduc represent formal call logic markov decis program lomdp integr markov decis process mdp logic program lomdp compactli declar repres complex mdp framework devis relat upgrad td call logic td prove converg balio engin bayesian logic program balio engin bayesian logic program probabilist logic learn past year wit signific interest probabilist logic learn research ly intersect probabilist reason logic represent machin learn rich varieti di erent formal learn techniqu develop paper introductori survei overview stateof art probabilist logic learn identif number import probabilist logic learn concept logic hidden markov model extend abstract kristian kerst tapani raiko stefan kramer luc de raedt machin learn lab comput scienc depart univers freiburg georg koehler all freiburg germani laboratori comput inform scienc helsinki univers technolog box helsinki fin hut finland keyword hidden markov model probabilist model order logic structur signatur scop grow number determin protein structur avail classif scheme increasingli import develop comput method automat extract structur signatur class protein present machin learn techniqu logic hidden markov model lohmm appli task find structur signatur fold classif scheme scop scale conjug gradient maximum likelihood empir comparison em algorithm abstract learn bayesian network estim paramet network data em expect maxim gradient base algorithm techniqu estim paramet theoret properti framework studi remain open question em prefer gradient answer question empir specif adapt scale conjug gradient neural network learn acceler conjug gradient avoid time consum line search tradit method empir compar scale conjug gradient em experi show acceler conjug gradient competit em gener em domain independ method choic gradient base method superior introduct bayesian network import effici eleg framework repres reason probabilist model joint probabl distribut finit set random variabl appli real world problem diagnosi forecast autom vision sensor fusion manufactur control past year interest problem learn bayesian network data avoid problem knowledg elicit learn bayesian network paramet estim fundament task inabl human reliabl estim paramet form basi learn problem adapt bayesian logic program order probabilist logic combin order logic probabilist knowledg represent context introduc continu bayesian logic program extend recent introduc bayesian logic program deal continu random variabl bayesian logic program tightli integr definit logic program bayesian network result framework nice seper qualit logic compon quantit probabilist show quantit compon learn gradient base maximum likelihood method combin induct logic program bayesian network recent represent languag integr order logic bayesian network develop bayesian logic program languag paper present result combin induct logic program ilp bayesian network learn qualit quantit compon bayesian logic program precis show combin ilp set learn interpret score base techniqu learn bayesian network paper posit answer koller pfeffer question techniqu ilp learn logic compon order probabilist model bayesian logic program propos combin order logic bayesian net exist introduc formal bayesian logic program basic simplif reformul ngo haddawi probabilist logic program bayesian logic program suffici power repres essenti knowledg eleg manner eleg illustr fact repres bayesian net definit claus program pure prolog kernel prolog adapt usual prolog meta interpret bayesian logic program propos combin order logic bayesian net exist introduc formal bayesian logic program basic simplif reformul ngo haddawi probabilist logic program bayesian logic program suffici power repres essenti knowledg eleg manner eleg illustr fact repres bayesian net definit claus program pure prolog kernel prolog adapt usual prolog meta interpret 