Value-Function Approximations for Partially Observable Markov Decision Processes