Neural network learning using the Robbins-Monro algorithm

Various learning algorithms for autoassociative Hopfield neural networks are compared. Such neural networks store a set of patterns by making the patterns stable states of the network. All of the algorithms studied (except one) are basically gradient-descent algorithms. They were compared by measuring how well they stored sets of randomly generated patterns. Each bit of each pattern generated was given an equal probability of being +1 or -1. After training the network for a set of patterns, the performance of the algorithm was tested by determining the next stable state for each possible input pattern and checking if that state was the closest in Hamming distance to the input pattern. The performance measurement is the percentage of input patterns that lead to the closest stable state