Markov decision processes under observability constraints

Abstract

We develop an algorithm to compute optimal policies for Markov decision processes subject to constraints that result from some observability restrictions on the process. We assume that the state of the Markov process is unobservable. There is an observable process related to the unobservable state. So, we want to find a decision rule depending only on this observable process. The objective is to minimize the expected average cost over an infinite horizon. We also analyze the possibility of performing observations in more detail to obtain improved policies.