Stochastic optimization of controlled partially observable Markov decision processes

We introduce an on-line algorithm for finding local maxima of the average reward in a Partially Observable Markov Decision Process (POMDP) controlled by a parameterized policy. Optimization is over the parameters of the policy. The algorithm's chief advantages are that it requires only a single sample path of the POMDP, it uses only one free parameter beta is an element of [0, 1), which has a natural interpretation in terms of a bias-variance trade-off, and it requires no knowledge of the underlying state. In addition, the algorithm can be applied to infinite state, control and observation spaces. We prove almost-sure convergence of our algorithm, and show how the correct setting of 0 is related to the mixing time of the Markov chain induced by the POMDP.