An approximate algorithm, with bounds, for composite-state partially observed Markov decision processes

The author presents an approximate algorithm with bounds, for solving composite-state POMDPs (partially observed Markov decision processes). The approximation is based on a discretization of the unit simplex that has proven effective with conventional POMDPs. The model considered is a composite-state space variation of the discrete-time, finite partially observed Markov decision process with stationary cost data analyzed by R.D. Smallwood and E.J. Sondik (1973). The computational savings achievable with the composite-state construction are indicated.<
>