Conversion of MDP problems into heuristics based planning problems using temporal decomposition

This paper presents an approach for recasting Markov Decision Process (MDP) problems into heuristics based planning problems. The basic idea is to use temporal decomposition of the state space based on a subset of state space referred to as termination sample space. Specifically, the recasting of MDP problems is done in three steps. First step is to define a state space adaptation criterion based on the termination sample space. Second step is to define an action selection heuristic from each state. Third and final step is to define a recursion or backtracking methodology to avoid dead ends and infinite loops. All three steps have been described and discussed. A case study involving fault detection and alarm generation for the reaction wheels of a satellite mission has been discussed. The proposed approach has been compared with existing approaches for recasting MDP problems using the case study. Computational reduction achieved by the proposed approach is evident from the results.