Structure determination of artificial neural network using modified cellular encoding

Abstract

This paper works with a new evolutionary system to construct and control the structure of feed forward artificial neural networks (ANNs), represented by modified cellular encoding (MCE) that is not subject to the well-known permutation problem. It is shown in this paper that addition or deletion of nodes or connections can evidently be done by crossover automatically. Hence, the number of user specified parameter is also decreased. The ANN architecture determination algorithm is tested on some real world problems. The algorithm is made adaptive.