Approximate stochastic annealing for online control of infinite horizon Markov decision processes

We present an online simulation-based algorithm called Approximate Stochastic Annealing (ASA) for solving infinite-horizon finite state-action space Markov decision processes (MDPs). The algorithm estimates the optimal policy by sampling at each iteration from a probability distribution function over the policy space, which is updated iteratively based on the Q-function estimates obtained via a recursion of Q-learning type. By exploiting a novel connection of ASA to the stochastic approximation method, we show that the sequence of distribution functions generated by the algorithm converges to a degenerated distribution that concentrates only on the optimal policy. Numerical examples are also provided to illustrate the algorithm.