NewExperiments onEnsembles ofMultilayer Feedforward for Classiflcation Problems
Carlos Hemandez-Espinosa, Joaquín Torres-Sospedra, Mercedes Femaindez-Redondo · 2005
Asshowninthebibliography, training anensemble ofnetworks isaninteresting waytoimprove theperformance withrespect toasingle network. However there areseveral methods toconstruct theensemble. Inthis paperwepresent somenew results ina comparison oftwenty different methods. We havetrained ensembles of3,9,20and40 networks toshowresults inawidespectrum ofvalues. The results showthattheimprovement inperformance above9 networks intheensemble depends onthemethodbutitis usually low.Also, thebestmethodfora ensemble of3 networks iscalled Decorrelated andusesapenalty termin theusualBackpropagation function todecorrelate the network outputs intheensemble. Forthecaseof9and20 networks thebestmethodisconservative boosting. And finally for40networks thebest method isCels.