Multi-Agent Reinforcement Learning Algorithm with Variable Optimistic-Pessimistic Criterion

A reinforcement learning algorithm for multi-agent systems based on variable Hurwicz's optimistic-pessimistic criterion is proposed. The formal proof of its convergence is given. Hurwicz's criterion allows to embed initial knowledge of how friendly the environment in which the agent is supposed to function will be. Thorough testing of the developed algorithm against well-known reinforcement learning algorithms has shown that in many cases its successful performance can be explained by its tendency to force the other agents to follow the policy which is more profitable for it. In addition the variability of Hurwicz's criterion allowed it to converge to best-response against opponents with stationary policies.