Efficient planning in large POMDPs through policy graph based factorized approximations
Partially observable Markov decision processes (POMDPs) are widely used for planning under uncertainty. In many applications, the huge size of the POMDP state space makes straightforward optimization of plans (policies) computationally intractable. To solve this, we introduce an efficient POMDP planning algorithm. Many current methods store the policy partly through a set of "value vectors" which is updated at each iteration by planning one step further; the size of such vectors follows the size of the state space, making computation intractable for large POMDPs. We store the policy as a graph only, which allows tractable approximations in each policy update step: for a state space described by several variables, we approximate beliefs over future states with factorized forms, minimizing Kullback-Leibler divergence to the non-factorized distributions. Our other speedup approximations include bounding potential rewards. We demonstrate the advantage of our method in several reinforcement learning problems, compared to four previous methods. Copyright 2010 Springer-Verlag. This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder. JoP belongs to AIRC; JaP to AIRC and HIIT. The work was supported by TEKES, PASCAL2, and Academy of Finland decision 123983.