Date of Award

Spring 2003

Project Type

Dissertation

Program or Major

Computer Science

Degree Name

Doctor of Philosophy

First Advisor

Eugene Freuder

Abstract

The focus of the thesis is on improving solving constraint satisfaction problems (CSPs) that change with certain conditions. This special class of problems, which we call conditional CSPs, has proved very useful in modeling important applications, such product configuration and design, and distributed software diagnosis and network management. The problem conditions model choices customers make to configure a product, or they are installation settings or actual observations of a running system that is monitored for diagnosis purpose.

The key, novel contribution of this thesis are two approaches for improving solving methods and the use of random conditional CSPs to evaluate the performance of these methods. With the first approach we propose new algorithms for solving conditional CSPs. These algorithms propagate problem constraints and conditions. The second approach explores the feasibility of reformulating the problem into a standard CSP and introduces new reformulation algorithms.

The implementation results have been evaluated experimentally. The experimental design has extensive test suites of randomly generated standard and conditional CSPs for which general problem parameters, such as density and satisfiability, were varied, as well as specialized parameters that characterize the representation of problem conditions.

The significance of the work lies in the advance of problem resolution for the class of conditional CSPs and the experimental analysis for the proposed new algorithms. The limited solving developments known in the literature of the class of conditional CSPs, a backtrack search algorithm tested on a handful of small problem examples, have been taken an important step further and aligned with efforts reported for standard and other special classes of CSPs.

Download

Share

COinS