Influence of training sample preprocessing in generalization accuracy of multilayer perceptron

Abstract

Summary form only given. In this paper the behavior of multilayer perceptron (backpropagation algorithm) generalization accuracy using different pre-processing methods of training sample is investigated. In the experiments, diverse techniques were used. These were separated in two groups: the first one contains those that select a subset of the original sample; the second one clusters techniques whose starting point is a group of codebook prototypes. The tests were carried our with real and artificial data, corresponding to different types of problems. Experimental results show that the combination of both types of procedures gives, in most cases, the best behavior, that is, when it executes an initial filtering with methods of the first group, and later a technique of the second group is applied.