StGA: An application of a genetic algorithm to stochastic learning automata
Masaharu Munetomi, Yoshiaki Takai, Yoshiharu Sato · Systems and Computers in Japan · 1996
Abstract Reinforcement learning algorithms such as stochastic learning automata are used to increase the probability that an action will succeed in a stochastic environment. Stochastic learning automata generally have the property of extremely slow convergence when they have a large number of feasible actions. This paper presents a novel reinforcement learning algorithm that employs a genetic algorithm to accelerate convergence. This learning algorithm, called a stochastic genetic algorithm (StGA) samples a small number of actions among all feasible ones. Procedures of stochastic learning automata and genetic operations are applied to a set of sampled actions in order to search effectively a feasible space of actions and find the best ones. Through theoretical investigations, a proof is given on convergence of the StGA by using an ε‐optimality of the stochastic learning automata. Moreover, empirical simulations demonstrate the effectiveness of the StGA when there is a large space of feasible actions.