Fault-tolerance inclusion in neural networks by a concurrent training algorithm

A learning algorithm, referred to as concurrent training, based on genetic algorithms for a neural network with connected modules is described. The algorithm does not require the knowledge of training sets for each module so that all modules can be trained concurrently. For an N module system, N separate pools of chromosomes are maintained and updated. The concurrent training algorithm is applied to train multilayered feedforward networks by considering each layer of connections to be a 1-layer network module. The algorithm is tested using the 4-bit parity problem and a linearly nonseparable classification problem. Experiment results are presented and the learning behavior and performance is analyzed