Reward revision for partially observed Markov decision processes

We present an algorithm for accelerating a successive approximation solution procedure for the infinite horizon, expected discounted total reward partially observed Markov decision process (POMDP). This algorithm proceeds by simultaneously constructing and solving a sequence of dynamic programs, each of which is more computationally attractive than the dynamic program associated with the original POMDP. The form of each of these dynamic programs is determined by the solution of the preceding dynamic program. A numerical study is presented that indicates the potentially significant computational value of this algorithm.