Abstract

A genetic algorithm is a technique designed to search large problem spaces using the Darwinian concepts of evolution. Solution representations are treated as living organisms. The procedure attempts to evolve increasingly superior solutions. As in natural genetics, however, there is no guarantee that the optimum organism will be produced.

One of the problems in producing optimal organisms in a genetic algorithm is the difficulty of premature convergence. Premature convergence occurs when the organisms converge in similarity to a pattern which is sub-optimal, but insufficient genetic material is present to continue the search beyond this sub-optimal level, called a local maximum.

The prevention of premature convergence of the organisms is crucial to the success of most genetic algorithms. In order to prevent such convergence, numerous operators have been developed and refined. All such operators, however, rely on the property of the underlying problem that the evaluation of individuals is a computationally inexpensive process.

In this paper, the design of genetic algorithms which intentionally converge rapidly is addressed. The design considerations are outlined, and the concept is applied to an NP-Complete problem, known as a Crozzle, which does not have an inexpensive evaluation function. This property would normally make the Crozzle unsuitable for processing by a genetic algorithm. It is shown that a rapidly converging genetic algorithm can successfully reduce the effective complexity of the problem.

Department(s)

Computer Science

Comments

This report is substantially the Ph.D. dissertation of the first author, completed May 1993.

Report Number

CSC-93-12

Document Type

Technical Report

Document Version

Final Version

File Type

text

Language(s)

English

Rights

© 1993 University of Missouri--Rolla, All rights reserved.

Publication Date

01 May 1993

Download

Share

 
COinS