Open Collections

Exploring partially observable Markov decision processes by exploting structure and heuristic information

University of British Columbia

This thesis is about chance and choice, or decisions under uncertainty. The desire for creating an intelligent agent performing rewarding tasks in a realistic world urges for working models to do sequential decision making and planning. In responding to this grand wish, decision-theoretic planning (DTP) has evolved from decision theory and control theory, and has been applied to planning in artificial intelligence. Recent interest has been directed toward Markov Decision Processes (MDPs) introduced from operations research. While fruitful results have been tapped from research in fully observable MDPs, partially observable MDPs (POMDPs) are still too difficult to solve as observation uncertainties are incorporated. Abstraction and approximation techniques become the focus. This research attempts to enhance POMDPs by applying A l techniques. In particular, we transform the linear POMDP constructs into a structured representation based on binary decision trees and Bayesian Networks to achieve compactness. A handful of tree-oriented operations is then developed to perform structural belief updates and value computation. Along ii with the structured representation, we explore the belief space with a heuristic online search approach, in which best-first search strategy with heuristic pruning is employed. Experimenting with a structured testbed domain reveals great potentials of exploiting structure and heuristics to empower POMDPs for more practical applications.

Thesis/Dissertation

application/pdf

2009-03-09

For non-commercial purposes only, such as research, private study and education. Additional conditions apply, see Terms of Use https://open.library.ubc.ca/terms_of_use.

http://hdl.handle.net/2429/5772

Computer Science

University of British Columbia

UBCV

DSpace