8/14/2007 - The paper on complete neural network induction using Darwinian evolution is now available online both in pdf format and html at:

https://www.gene-expression-programming.com/webpapers/abstracts.asp#14

Ferreira, C., Designing Neural Networks Using Gene Expression Programming. In A. Abraham, B. de Baets, M. Köppen, and B. Nickolay, eds., Applied Soft Computing Technologies: The Challenge of Complexity, pages 517-536, Springer-Verlag, 2006.

ABSTRACT: An artificial neural network with all its elements is a rather complex structure, not easily constructed and/or trained to perform a particular task. Consequently, several researchers used genetic algorithms to evolve partial aspects of neural networks, such as the weights, the thresholds, and the network architecture. Indeed, over the last decade many systems have been developed that perform total network induction. In this work it is shown how the chromosomes of Gene Expression Programming can be modified so that a complete neural network, including the architecture, the weights and thresholds, could be totally encoded in a linear chromosome. It is also shown how this chromosomal organization allows the training/adaptation of the network using the evolutionary mechanisms of selection and modification, thus providing an approach to the automatic design of neural networks. The workings and performance of this new algorithm are tested on the 6-multiplexer and on the classical exclusive-or problems.

This paper requires a certain familiarity with the basics of GEP, especially the head/tail organization, the expression of genes with random constants, and the type and mechanisms of the genetic operators. For a quick introduction see my Complex Systems paper.

For the sample problems of this paper I chose well-known logical functions, but the beauty of GEP-nets is that they can be used on a multitude of modeling problems, from nonlinear regression to classification and they are as good as any GEP system. I guess I’ll have to write a paper on this since I haven’t seen anyone taking up on this task since I first described this algorithm in my 2002 book.