Created by W.Langdon from gp-bibliography.bib Revision:1.3872

@PhdThesis{Chellapilla:thesis, author = "Kumar H. Chellapilla", title = "Designing Effective Evolutionary Computations", school = "Electrical Engineering, University of California, San Diego", year = "2005", address = "USA", keywords = "genetic algorithms, genetic programming", isbn13 = "9780542572180", URL = "http://phdtree.org/pdf/26271964-designing-effective-evolutionary-computations/", URL = "http://search.proquest.com/docview/305003505", size = "606 pages", abstract = "Evolutionary algorithms offer a practical approach to solving difficult real-world problems. In many problem domains, these are the only possible approaches with potential for effectively searching through complex solution spaces. For novel problem domains wherein previous research efforts are sparse or problem domain expertise is in its infancy, evolutionary algorithms offer strong alternatives for exploring the solution spaces and also gaining insights into effectively solving the problem. The principal roadblock in conventional practice is the lack of a specific approach which permits one to simultaneously control an algorithm's representation, population variation operators and population selection operators. An approach based on mathematically sound principles is adopted in this thesis to provide asymptotic guarantees on evolutionary algorithm performance followed by useful real-time methods for improving the rate of convergence. In particular, the evolutionary algorithm is decomposed into its constituent representation, population variation, and population selection operators. The population variation operators are further broken down into solution variation operators. Each component is independently analysed without being constrained by an overall architecture for the evolutionary algorithm. Each component presents several alternatives that can be chosen independently to control desired properties of the evolutionary algorithm. A new mathematical model for analysing evolutionary algorithms is developed, and necessary and sufficient conditions on the variation and selection operators for asymptotic convergence are derived. Fitness distributions and fitness distribution feature based heuristics are presented to improve the rate of convergence of an evolutionary algorithm. This thesis also presents a wide array of empirical results to demonstrate the utility, effectiveness, and applicability of the new theory. Within the new framework, evolutionary algorithms are applied to solve real, discrete and mixed parameter optimization problems. Evolutionary algorithms that guarantee asymptotic convergence are designed to solve problems involving structures such as parse trees and finite state machines. Co-evolutionary algorithms are designed to evolve an expert checkers player that rated 2045 against human checkers players. Fitness distribution heuristics are used to tune an evolutionary algorithm for improved rate of convergence for solving the travelling salesman problem.", notes = "Supervisor Anthony Sebald UMI Microform 3208643", }

Genetic Programming entries for Kumar Chellapilla