A Scheme for the Evolution of Feedforward Neural Networks using BNF-Grammar Driven Genetic Programming

This paper presents our attempt to automatically define feedforward neural networks using genetic programming. Neural networks have been recognized as powerful approximation and classification tools. On the other hand, the genetic programming has been used effectively for the production of intelligent systems, such as the neural networks. In order to reduce the search space and guide the search process we employ grammar restrictions to the genetic programming population individuals. To implement these restrictions, we selected to apply a context-free grammar, such as a BNF grammar. The proposed grammar extends developments of cellular encoding, inherits present advances and manages to express arbitrarily large and connected neural networks. Our implementation uses parameter passing by reference in order to emulate the parallel processing of neural networks into the genetic programming tree individuals. The system is tested in two real-world domains denoting its potential future use.