AAAI Publications, Sixth European Conference on Planning

Font Size: 
Solving Informative Partially Observable Markov Decision Processes
Nevin L. Zhang, Weihong Zhang

Last modified: 2014-05-21


Solving Partially Observable Markov Decision Processes (POMDPs) generally is computationally intractable. In this paper, we study a special POMDP class, namely informative POMDPs, where each observation provides good albeit incomplete information about world states. We propose two ways to accelerate value iteration algorithm for such POMDPs. First, dynamic programming (DP) updates can be carried out over a relatively small subset of belief space. Conducting DP updates over subspace leads to two advantages: representational savings in space and computational savings in time. Second, a point-based procedure is used to cut down the number of iterations for value iteration over subspace to converge. Empirical studies are presented to demonstrate various computational gains.


POMDPs, value iteration, point-based updating, acceleration

Full Text: PDF