Evolving Modular Neural Networks Using Rule-Based Genetic Programming

摘要

This paper describes a new approach for evolving recurrent neural networks using Genetic Programming. A system has been developed to train weightless neural networks using construction rules. The network construction rules are evolved by the Genetic Programming system which build the solution neural networks. The use of rules allows networks to be constructed modularly. Experimentation with decomposable Boolean functions has revealed that the performance of the system is superior to a non-modular version of the system.