Evolving neural networks for decomposable problems using genetic programming

Many traditional methods for training neural networks using genetic algorithms and genetic programming do not have any special provisions for taking advantage of decomposable problems which can be solved by combining solutions to each subproblem. This paper describes a new approach to neural network construction using genetic programming which is designed to rapidly construct networks composed of similar subnetworks. A system has been developed to produce trained weightless neural networks by using construction rules to build the networks. The network construction rules are evolved by the genetic programming system. The system has been applied to decomposable Boolean problems and the results were compared with a modified version of the system in which networks cannot be constructed modularly. The modular version of the system obtains significantly better results than the nonmodular version of the program.