Sample path average optimality of Markov control processes with strictly unbounded cost

Abstract

We study the existence of sample path average cost (SPAC-) optimal policies for Markov control processes on Borel spaces with strictly unbounded costs, i.e., costs that grow without bound on the complement of compact subsets. Assuming only that the cost function is lower semicontinuous and that the transition law is weakly continuous, we show the existence of a relaxed policy with “minimal” expected average cost and that the optimal average cost is the limit of discounted programs. Moreover, we show that if such a policy induces a positive Harris recurrent Markov chain, then it is also sample path average (SPAC-) optimal. We apply our results to inventory systems and, in a particular case, we compute explicitly a deterministic stationary SPAC-optimal policy.