Finite horizon partially observable semi-Markov decision processes under risk probability criteria

This paper deals with a risk probability minimization problem for finite horizon partially observable semi-Markov decision processes, which are the fairly most general models for stochastic dynamic systems. In contrast to the expected discounted and average criteria, the optimality investigated in this paper is to minimize the probability that the accumulated rewards do not reach a prescribed profit level at the finite terminal stage. First, the state space is augmented as the joint conditional distribution of the current unobserved state and the remaining profit goal. We introduce a class of policies depending on observable histories and a class of Markov policies including observable process with the joint conditional distribution. Then under mild assumptions, we prove that the value function is the unique solution to the optimality equation for the probability criterion by using iteration techniques. The existence of (ϵ-)optimal Markov policy for this problem is established. Finally, we use a bandit problem with the probability criterion to demonstrate our main results in which an effective algorithm and the corresponding numerical calculation are given for the semi-Markov model. Moreover, for the case of reduction to the discrete-time Markov model, we derive a concise solution.