Probabilistic planning via automatic feature induction and stochastic enforced hill-climbing

Jia-Hong Wu, Purdue University

Abstract

State-value functions are often critical to the success of decision-making processes. Such value functions are used to estimate the importance of world states or the distance to some desired goal regions. Value functions may be used as search heuristics in decision-theoretic planning. In the first part of this dissertation, we present a framework that finds useful value functions in stochastic planning via automated feature induction. We use inconsistencies found in Bellman equation to form “ideal” Bellman error features, and apply machine learning methods to find compact approximations of such Bellman error features. We utilize both relational and propositional feature representations, and evaluate the performance of the learned features using a collection of stochastic planning domains. We show that our method is the state-of-the-art among feature-learning stochastic planners, and also performs significantly better than a previous competition winner in several domains. In the second part of this dissertation, we consider how to utilize useful but flawed value functions as search heuristics in stochastic planning. We propose stochastic enforced hill-climbing (SEH), a stochastic local search technique. SEH is a stochastic generalization of a proven deterministic planning technique, enforced hill-climbing, that uses breadth-first search to escape local optimality caused by flawed heuristics. In SEH, we find local policies with expected improvement on the heuristic value of the current state. SEH is evaluated using the planning competition domains for fully-observable probabilistic planners and is shown to be the state-of-the-art.

Degree

Ph.D.

Advisors

Givan, Purdue University.

Subject Area

Computer science

Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server
.

Share

COinS