Exact Dynamic Programming for Decentralized POMDPs with Lossless Policy Compression

High dimensionality of belief space in DEC-POMDPs is one of the major causes that makes the optimal joint policy computation intractable. The belief state for a given agent is a probability distribution over the system states and the policies of other agents. Belief compression is an efficient POMDP approach that speeds up planning algorithms by projecting the belief state space to a low-dimensional one. In this paper, we introduce a new method for solving DEC-POMDP problems, based on the compression of the policy belief space. The reduced policy space contains sequences of actions and observations that are linearly independent. We tested our approach on two benchmark problems, and the preliminary results confirm that Dynamic Programming algorithm scales up better when the policy belief is compressed