Adaptive Representations for Genetic Algorithms and Local Search

Colorado State University, Fort Collins CO

Investigators

Abstract

Adaptive Representations for Genetic Algorithms and Local Search This is the first year funding of a three year continuing award. The PI's recent theoretical results prove there are advantages to reflected Gray codes compared to standard binary encodings when used as a representation for search and parameter optimization problems (e.g., it is possible to use a simple mechanism to escape local optima by adaptively switching between different reflected Gray codes). Preliminary empirical results indicate that using high precision adaptive Gray codes leads to better optimization, and new convergence results have been proven. In this project, a family of high precision adaptive Gray codes will be developed for use with genetic algorithms and local search. The PI will further conduct a broad comparative evaluation of evolutionary algorithms, local search and other heuristic search methods that do not require derivative information. A real-world application will be addressed related to a new generation of much more accurate weather prediction systems. The CloudSat project will in the next few years deploy a satellite-based system for modeling cloud structures, and the search methods developed in this project will be the best and fastest way to solve a key inverse problem which lies at the heart of such models, as a consequence of which they appear to be likely candidates for inclusion in the scheduled 2003 deployment of CloudSat.