Optimistic-Pessimistic Q-Learning Algorithm for Multi-Agent Systems

摘要

A reinforcement learning algorithm OP-Q for multi-agent systems based on Hurwicz's optimistic-pessimistic criterion which allows to embed preliminary knowledge on the degree of environment friendliness is proposed. The proof of its convergence to stationary policy is given. Thorough testing of the developed algorithm against well-known reinforcement learning algorithms has shown that OP-Q can function on the level of its opponents.