Genetic Encoding of Neural Networks using Attribute Grammars

Talib S. Hussain and Roger A. Browse

Abstract

The discovery of good neural network solutions to complex problems may be facilitated through the use of evolutionary computation techniques, such as genetic algorithms or genetic programming. One key issue in the development of any system which will evolve neural networks is how and what information about a neural network will be encoded in the genetic description that will be manipulated by the evolutionary processes. Several approaches have been taken to this encoding problem, including direct, structural, parametric, and grammatical encoding. We present a new grammatical encoding technique in which an attribute grammar is used to represent a class of neural networks. We propose that the resulting encoding offers several improvements over existing approaches.