Decentralized learning in finite Markov chains

The principal contribution of this paper is a new result on the decentralized control of finite Markov chains with unknown transition probabilities and rewards. One decentralized decision maker is associated with each state in which two or more actions (decisions) are available. Each decision maker uses a simple learning scheme, requiring minimal information, to update its action choice. It is shown that, if updating is done in sufficiently small steps, the group will converge to the policy that maximizes the long-term expected reward per step. The analysis is based on learning in sequential stochastic games and on certain properties, derived in this paper, of ergodic Markov chains.