Hierarchical asynchronous genetic algorithms for parallel/distributed simulation-based optimization.

The objective of this dissertation is to develop a multi-resolution optimization strategy based on the evolution algorithms in the parallel/distributed simulation environment. The system architecture is constructed hierarchically with multiple clusters which consist of an expert system (controller) and set of genetic algorithm optimizers (agents). We propose an asynchronous genetic algorithm (AGA) which continuously updates the population in parallel genetic algorithms. Asynchronous evaluation of population in a parallel computer improves the utilization of the processors and reduces search time when the evaluation time of individuals is highly variable. Further, we have devised a noise assignment scheme which resolves the pre-convergence drawback of the genetic algorithms. In this scheme, binary representation (discrete sampling) of an individual is combined with a random number (analog sampling), so that the genetic algorithm can investigate the entire search space regardless of the bit-size of an individual. Real application problems require the evaluation of a large number of parameters and their search complexity grows beyond the capability of a single level GA-optimizer. In response, we have developed a novel scheme called Hierarchical Genetic Algorithms. This multilevel GA optimization strategy is based on an Intelligent Machine Architecture which supporting non-deterministic computation, intensive and irregular memory access patterns, and large potential for parallel computing. The clusters in the Hierarchical GAs are coordinated hierarchically and creation and deletion of nodes are performed dynamically based upon performance. During the optimization process, the clusters cooperate together to solve different levels of the abstracted problem. A candidate solution at a higher level creates a lower level cluster which utilizes previously optimized parameter information. It can also contribute to the search process of a higher level by sending the feedback information. Hierarchical GAs demonstrate performance with various experiments. We have compared the performance of the Hierarchical GAs and simple GA (single-level). The Hierarchical GAs adaptively changes its structure to allocate more computing resources to the promising nodes. With the same amount of memory size for population, the simulation results shows that the Hierarchical GAs find a solution faster than the simple GA.This item was digitized from a paper original and/or a microfilm copy. If you need higher-resolution images for any content in this item, please contact us at repository@u.library.arizona.edu.Original file replaced with corrected file November 2023