Feature Selection Based onGjenet'ic Alor'tithms forOn-Line Signature Vterification
Javier Galbally, Julián Fiérrez · 2007
error rate ofthesystem. Thecurse ofdimensionality phenomenon solution inthecomplete space of2100possible solutions. isfurther investigated using aGA withinteger coding. Results * GA withinteger coding: itsearches forasuboptimal aregiven ontheMCYT signature database comprising 330users solution ofa specified dimension d.Inthiscasethe (16500 signatures). Signatures arerepresented bymeansofaset 100 of100features whichcanbedivided intofourdifferent groups dimension ofthesearch space is d10 according tothesignature information theycontain, namely: i) d time, ii)speed andacceleration, iii) direction, andiv)geometry.Fourdifferent scenarios areonsidered: skilled andrandom TheGA indicates thatfeatures fromsubsets iandivarethe forgeries with5 and20training signatures. Theoriginal mostdiscriminative whendealing withrandomforgeries, whilefeatures aredivided into fourdifferent groups according tothe parameters fromsubsets iiandivarethemostappropriate to signature information theycontain, namely: i)time, ii)speed maximize therecognition ratewithskilled forLgeries. miehegireislf i andacceleration, i)direction, andv)geometry. Comparative