Associative Reinforcement Learning of Real-valued Functions

|Associative reinforcement learning (ARL) tasks de ned originally by Barto and Anandan [1] combine elements of problems involving optimization under uncertainty, studied by learning automata theorists, and supervised learning pattern-classi cation. The stochastic real-valued (SRV) unit algorithm [6] has been designed for an extended version of ARL tasks wherein the learning system's outputs can take on real values. In this paper, we present a strong convergence theorem that implies a form of optimal performance (under certain general conditions) of the SRV algorithm on ARL tasks. Simulation results are presented to illustrate the convergence behavior of the algorithm under the conditions of the theorem. The robustness of the algorithm is also demonstrated by simulations in which some of the conditions of the theorem are violated.