Date

2013

Document Type

Dissertation

Degree

Doctor of Philosophy

Department

Computer Science

First Adviser

Munoz-Avila, Hector

Other advisers/committee members

Kay, Edwin; Heflin, Jeff; O'Seaghdha, Padraig

Abstract

Planning is a branch of Artificial Intelligence (AI) concerned with projecting courses of actions for executing tasks and reaching goals. AI Planning helps increase the autonomy of artificially intelligent agents and decrease the cognitive load burdening human planners working in challenging domains, such as the Mars exploration projects. Approaches to AI planning include first-principles heuristic search planning and case-based planning. The former conducts a heuristic-guided search in the solution space, while the latter generates new solutions by adapting solutions to previously-solved problems.The ability to generate not just one solution, but a set of meaningfully diverse solutions to each planning problem helps cater to a wider variety of user preferences and needs (which it may be difficult or even unfeasible to acquire and/or represent in their entirety), produce viable alternative courses of action to fall back on in case of failure, counter varied threats in intrusion detection, render computer games more compelling, and provide representative samples of the vast search spaces of planning problems.This work describes a general framework for generating diverse sets of solutions (i.e. courses of action) to planning problems. The general diversity-aware planning algorithm consists of iteratively generating solutions using a composite candidate-solution evaluation criterion taking into account both how promising the candidate solutions appear in their own right and on how likely they are to increase the overall diversity of the final set of solutions. This estimate of diversity is based on distance metrics, i.e. measures of the dissimilarity between two solutions. Distance metrics can be quantitative or qualitative.Quantitative distance measures are domain-independent. They require minimum knowledge engineering, but may not reflect dissimilarities that are truly meaningful. Qualitative distance metrics are domain-specific and reflect, based on the domain knowledge encoded within them, the kind of meaningful dissimilarities that might be identified by a person familiar with the domain.Based on the general framework for diversity-aware planning, three domain-independent planning algorithms have been implemented and are described and evaluated herein. DivFF is a diverse heuristic search planner for deterministic planning domains (i.e. domains for which the assumption is made that any action can only have one possible outcome). DivCBP is a diverse case-based planner, also for deterministic planning domains. DivNDP is a heuristic search planner for nondeterministic planning domains (i.e. domains the descriptions of which include actions with multiple possible outcomes). The experimental evaluation of the three algorithms is conducted on a computer game domain, chosen for its challenging characteristics, which include nondeterminism and dynamism. The generated courses of action are run in the game in order to ascertain whether they affect the game environment in diverse ways. This constitutes the test of their genuine diversity, which cannot be evaluated accurately based solely on their low-level structure.It is shown that all proposed planning systems successfully generate sets of diverse solutions using varied criteria for assessing solution dissimilarity. Qualitatively-diverse solution sets are demonstrated to constantly produce more diverse effects in the game environment than quantitatively-diverse solution sets.A comparison between the two planning systems for deterministic domains, DivCBP and DivFF, reveals the former to be more successful at consistently generating diverse sets of solutions. The reasons for this are investigated, thus contributing to the literature of comparative studies of first-principles and case-based planning approaches. Finally, an application of diversity in planning is showcased: simulating personality-trait variation in computer game characters. Sets of diverse solutions to both deterministic and nondeterministic planning problems are shown to successfully create diverse character behavior in the evaluation environment.

Share

COinS