Clustered Recursive Branching Network

摘要

Recursive branching network (RBN) was proposed in [1] to solve linearly non-separable problems using output-coded perceptrons. It relies on splitting the training patterns, at random, between parallel perceptrons. However, the random splitting mechanism can trap the perceptron with conflicting patterns. Optimizing the splitting methods, through clustering, is proposed here to ensure meaningful way of distributing the patterns between the perceptrons. We propose four splitting methods which use different similarity measures between patterns. We examine these methods on five standard data sets. In general, these methods enhance the performance of RBN and in many cases contribute to reducing the network complexity compared with random-splitting RBN.