"On Combining Artificial Neural Nets: Ensemble approach"

Supervised by A.Sharkey

Abstract

Several researchers have experimentally shown that substantial improvements can be obtained in pattern recognition problems by combining the outputs of multiple classifiers. The use of such a technique, often called the ensemble approach to combining predictors, has been shown not to require expensive supplementary computation to when techniques for creating accurate unitary classifiers are employed.

An investigation of two major aspects of the ensemble approach, member creation and combination, is undertaken. The presence of coincident errors between ensemble members is established to be the main restriction to the improvement of ensemble performance. It is seen to be likely that a population of neural networks has coincident errors. During an investigation of methods to minimise the effect these errors have on the aggregated output, a new procedure for ensemble member selection is proposed and evaluated along with a new variation of the majority vote scheme. The two new procedures are tested against other similar methods; the new methods were seen to perform better.