Global convergence of feedforward networks of learning automata

A feedforward network composed of units of teams of parameterized learning automata is considered as a model of a reinforcement learning system. The parameters of each learning automaton are updated using an algorithm consisting of a gradient following term and a random perturbation term. The algorithm is approximated by the Langevin equation. It is shown that it converges to the global maximum. The algorithm is decentralized and the units do not have any information exchange during updating. Simulation results on a pattern recognition problem show that reasonable rates of convergence can be obtained.<>