Multilayer Networks of Self-Interested Adaptive Units

摘要

This report describes research directed toward refining and evaluating learning methods for multilayer networks of neuron-like adaptive units. We define a learning rule called the Associative Reward-Penalty, or A sub R-P, rule that has strong ties to both the theory of adaptive pattern classification and stochastic learning automata. We state a convergence result that has been proven for a single A sub R-P units can reliably learn nonlinear associative mappings. The behavior of these networks is discussed in terms of the collective behavior of stochastic learning automata in team decision problems. A number of methods for learning in multilayer networks are compared, including the A sub R-P method and the error back-propagation method. These methods, or variants of them, outperform the other methods applied to the test problem, with error back-propagation showing a significant speed advantage over the other methods. The A sub R-P and error back-propagation are compared and contrasted in terms of their respective approaches to gradient following.