Genetic Design of Sparse Feedforward Neural Networks

Abstract A novel scheme has been proposed to design neural networks. The method attempts to find an architecture and the corresponding weights and thresholds of the neural network simultaneously. Various rules have been formulated to encode the architecture and network parameters into a string of bits, and the genetic algorithm has been used to find a beneficial (if not optimal) string. The proposed scheme has been tested on the XOR problem. The neural networks which have been designed by this approach include interesting examples of shortcut or bypass, and sparse neural networks.