Evaluating Effects of Two Alternative Filters for the Incremental Pruning Algorithm on Quality of Pomdp Exact Solutions

摘要

Decision making is one of the central problems in artificial intelligence and specifically in robotics. In most cases this problem comes with uncertainty both in data received by the decision maker/agent and in the actions performed in the environment. One effective method to solve this problem is to model the environment and the agent as a Partially Observable Markov Decision Process (POMDP). A POMDP has a wide range of applications such as: Machine Vision, Marketing, Network troubleshooting, Medical diagnosis etc. In recent years, there has been a significant interest in developing techniques for finding policies for (POMDPs).We consider two new techniques, called Recursive Point Filter (RPF) and Scan Line Filter (SCF) based on Incremental Pruning (IP) POMDP solver to introduce an alternative method to Linear Programming (LP) filter for IP. Both, RPF and SCF have solutions for several POMDP problems that LP could not converge to in 24 hours. Experiments are run on problems from POMDP literature, and an Average Discounted Reward (ADR) is computed by testing the policy in a simulated environment.