Biologically influenced algorithms and parallelism in non-linear optimization

This dissertation examines the use of the Genetic Algorithm (GA) as implemented on the Connection Machine Model 2 (CM-2) for use in solving difficult problems in function optimization. A variety of panmictic mating strategies are compared with geographically-structured mating strategies in terms of population convergence, inbreeding, and fitness. Various parallel GA implementations for both large-grained and fine-grained parallel architectures are surveyed. Issues from population genetics, including Sewall Wright's shifting balance theory of evolution as implemented with a diffusion model, are explored as the mechanism for the power inherent in this model. Several difficult function optimization problems are evaluated including Kan's function in 40 dimensions which, in this model, required approximately 6.3 $\times$ 10$\sp9$ function evaluations and over 9 hours of CM-2 time using 16K processors.