cohen-icml95 - T oappearin/M ach ineL earn ing :P roceed...

Info iconThis preview shows pages 1–2. Sign up to view the full content.

View Full Document Right Arrow Icon

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: T oappearin/M ach ineL earn ing :P roceed ingso ftheTw e lfthIn ternat iona lC on ference"(M L) F astEect iveR u leInduct ion W illiamW .C ohen AT&TB e llLaborator ies M ounta inA venueM u rrayH ill,N J [email protected] A b stract M anyex ist ingru lelearn ingsy stem sare com pu tat iona llyexpen s iveonlargeno isy datasets . Inth ispaperw eeva luatethe recen t ly-p roposedru lelearn inga lgor ithm IREPonalargeandd iverseco llect iono f benchm arkp rob lem s .W eshow thatwh ile IREP isex trem e lyec ient ,itfrequen t ly g iveserrorratesh igherthanthoseo fC . andC .ru les . W ethenp roposeanum- bero fm od icat ion sresu lt inginana lgo- r ithmR IPPER k thatisverycom pet it ive w ithC .ru lesw ithrespecttoerrorrates , bu tm uchm oreec ientonlargesam p les . R IPPER k ob ta in serrorrateslow erthanor equ iva len ttoC .ru leson o f bench- m arkp rob lem s ,sca lesnear lylinear lyw ith thenum bero ftra in ingexam p les ,andcan ec ient lyp rocessno isydatasetscon ta in ing hund red so fthou sand so fexam p les . INTRODUCT ION Sy stem sthatlearnsetso fru leshaveanum bero fde- s irab lep ropert ies . R u lesetsarere lat ive lyeasyfor peop letounderstand [C at lett , ],andru lelearn- ingsy stem sou tperform dec is iontreelearnersonm any p rob lem s[P aga lloandH au ss ler , ;Q u in lan , ; W e issandIndu rkhya , ].R u lesetshaveanat- u ra landfam iliarrstordervers ion ,nam e lyP ro log p red icates ,andtechn iquesforlearn ingp ropos it iona l ru lesetscano ftenbeex tendedtotherst-ordercase [Q u in lan , ; Q u in lanandC am eron-Jones , ]. C erta intypeso fp r iorknow ledgecana lsobeeas ily comm un icatedtoru lelearn ingsy stem s[C ohen , ; P azzan iandK ib ler , ]. O new eaknessw ithru lelearn ingsy stem sisthatthey o ftensca lere lat ive lypoor lyw iththesam p les ize ,par- t icu lar lyonno isydata [C ohen , ].G iventhep reva- lenceo flargeno isydatasetsinrea l-w or ldapp licat ion s , th isp rob lem iso fcr it ica lim portance .Thegoa lo fth is paperistodeve lopp ropos it iona lru lelearn inga lgo- r ithm sthatperform ec ient lyon largeno isydatasets , thatex tendnatu ra llytorst-orderrep resen tat ion s , andthatarecom pet it iveingenera lizat ionperform ance w ithm orem atu resym bo liclearn ingm ethod s ,such asdec is iontrees .Theendp roducto fth iseortis thea lgor ithmR IPPER k ,wh ich iscom pet it ivew ith C .ru lesw ithrespecttoerrorrates ,sca lesnear ly lin- ear lyw iththenum bero ftra in ingexam p les ,andcan ec ient lyp rocessno isydatasetscon ta in inghund red s o fthou sand so fexam p les . PREV IOU SW ORK . COM PLEX ITYOFRULEPRUN ING M anyo fthetechn iquesu sed inm odernru lelearn- ershavebeenadap tedfromdec is iontreelearn ing . M ostw ide ly-u seddec is iontreelearn ingsy stem su sean overt-and-sim p lify learn ingstrategytohand leno isy data :ahypothes isisform edbyrstgrow ingacom- p lextreewh ich/overts"thedata ,andthens im- p lify ingor prun ing thecom p lextree[Q u in lan , ; M ingers , ].U sua lly(bu tnota lw ay s)suchp run ing...
View Full Document

This note was uploaded on 02/10/2012 for the course CSE 5800 taught by Professor Staff during the Fall '09 term at FIT.

Page1 / 9

cohen-icml95 - T oappearin/M ach ineL earn ing :P roceed...

This preview shows document pages 1 - 2. Sign up to view the full document.

View Full Document Right Arrow Icon
Ask a homework question - tutors are online