Including control architecture in attribute grammar specifications of feedforward neural networks

Talib S. Hussain and Roger A. Browse

Abstract

An important problem in evolutionary computing is the design of genetic representations of neural networks that permit optimization of topology and learning characteristics. One promising approach for genetic representation of neural networks is the use of grammars to depict a process in which neural networks may be generated. Existing grammar representations of neural networks describe classes of networks with homogenous processing elements, simple fixed learning mechanisms and little organized topological structure. In previous research we have presented an attribute grammar representation for classes of networks with modular topology. Each parse tree generated by the grammar encodes a neural network specification which is subsequently executed by an interpreter. By expanding the grammar to include the control of the sequence of activity in the networks, we have been able to reduce the interpreter to a simple model of the operation of individual neurons in the networks. The expanded grammar may be used in an evolutionary computation to explore neural networks that vary in their architecture and learning behaviors.