MSc Graduation Seminar - Measurement Simplification in POMDP with Performance Guarantees

Advisor: Associate Professor Vadim Indelman

Abstract:

POMDP planning is at the heart of any autonomous system acting with imperfect information. The cost of solving the POMDP planning problem is exponential in the action and observation spaces, thus rendering it unfeasible for many online systems. We propose a novel approach to efficient decision making, by simplifying the observation space. We formulate analytical bounds on the expected information-theoretic reward, for the most general belief distributions. These bounds are then used to plan efficiently while keeping performance guarantees. We propose a specific variant of these bounds for Gaussian beliefs and show performance improvement of at least a factor of 4.