Date of Award

Spring 2012

Project Type

Dissertation

Program or Major

Computer Science

Degree Name

Doctor of Philosophy

First Advisor

Wheeler Ruml

Abstract

Intelligence is difficult to formally define, but one of its hallmarks is the ability find a solution to a novel problem. Therefore it makes good sense that heuristic search is a foundational topic in artificial intelligence. In this context "search" refers to the process of finding a solution to the problem by considering a large, possibly infinite, set of potential plans of action. "Heuristic" refers to a rule of thumb or a guiding, if not always accurate, principle. Heuristic search describes a family of techniques which consider members of the set of potential plans of action in turn, as determined by the heuristic, until a suitable solution to the problem is discovered.

This work is concerned primarily with suboptimal heuristic search algorithms. These algorithms are not inherently flawed, but they are suboptimal in the sense that the plans that they return may be more expensive than a least cost, or optimal, plan for the problem. While suboptimal heuristic search algorithms may not return least cost solutions to the problem, they are often far faster than their optimal counterparts, making them more attractive for many applications.

The thesis of this dissertation is that the performance of suboptimal search algorithms can be improved by taking advantage of information that, while widely available, has been overlooked. In particular, we will see how estimates of the length of a plan, estimates of plan cost that do not err on the side of caution, and measurements of the accuracy of our estimators can be used to improve the performance of suboptimal heuristic search algorithms.

Download

Share

COinS