Improving classification performance in the bumptree network by optimising topology with a genetic algorithm

This paper presents a successful synthesis of evolutionary and connectionist methods, based on the genetic optimisation of a recently introduced neural network model, the bumptree network. We show that the bumptree network is inherently more suited to optimisation by a genetic algorithm (GA) than other neural network models such as the multi-layer perceptron (MLP). We describe a hierarchical genetic coding which addresses the problem of representing certain strong dependencies which exist between the bumptree´s structural parameters, and show that our coding scheme has the desirable properties of continuity, isomorphism, completeness, closure and low redundancy with respect to the space of possible bumptree structures. We present empirical results which show that bumptree networks evolved by the GA significantly outperform the orthodox bumptree on several tasks, including the difficult real-world classification task of spoken vowel recognition