A Fast Approximation Method for Partially Observable Markov Decision Processes

摘要

This paper develops a new lower bound method for POMDPs that approximates the update of a belief by the update of its non-zero states.It uses the underlying MDP to explore the optimal reachable state space from initial belief and select actions during value iterations,which significantly accelerates the convergence speed.Also,an algorithm which collects and prunes belief points based on the upper and lower bounds is presented,and experimental results show that it outperforms some of the state-of-art point-based algorithms.