The Optimal Control of Partially Observable Markov Processes over a Finite Horizon