NOISY REINFORCEMENT TRAINING FOR PRAM NETS

The use of additional noise in reinforcement training of probabilistic RAMS (pRAMs) is analysed in the context of pattern recognition. Both simulations and analysis indicate the effectiveness of adding a contro lled level of noise during training. If the characteristics of the add ed noise match the noise expected in the testing signals then optimal behaviour is to be expected. It is shown how noise broadens the basins of attraction for the net and this is achieved in a manner determined by the characteristics of the training set, not by a separate general isation process. Mathematical analysis of the asymplotic values of the weights is performed and is shown to agree with the results obtained by simulation. This analysis is extended to multiple layers. Finally, an analysis of the weights in a trained net is performed to illustrate how training noise both forms the basins of attraction and also achie ves the maximum distance between attractors for optimum classification performance.