Robust learning algorithms for multi-layer perceptrons with discretized synaptic weights

The multi layer perceptron is one of the most popular artificial neural networks with applications to e.g. signal and image processing as well as pattern recognition and classification. In order to implement this network and the corresponding learning algorithm in electronic or optoelectronic hardware, the synaptic weights have to be discretized. This discretization, however, poses difficulties for the successful training of a multi layer perceptron. Variants of the FLETCHER-REEVES algorithm are compared to the genetic algorithm as learning strategies for the multi layer perceptron with discretized synaptic weights. Simulation results with the XOR problem as benchmark are given.