Decentralized learning automata for control of unknown markov chains

In this paper, we propose a new type of decentralized learning automata for the control finite state Markov chains with unknown transition probabilities and rewards. In our scheme a .betha.-type learning automaton is associated with each state in which two or more actions(desisions) are available. In this decentralized learning automata system, each learning automaton operates, requiring only local information, to improve its performance under local environment. From simulation results, it is shown that the decentralized learning automata will converge to the optimal policy that produces the most highly total expected reward with discounting in all initiall states.