A study of mate selection in genetic algorithms.

The process of information exchange among the population of individuals manipulated by Genetic Algorithms (GAs) involves two key components: crossover and mate selection. The central theme of this thesis concentrates on the investigation of effects of mate selection in GAs. The importance of mate selection in biology is widely recognized, yet a systematic investigation of this subject in GA research is still lacking. The goal of this thesis is to propose a framework that facilitates exploration of mate selection in GAs in order to (1) gain a deeper understanding of how GAs work, (2) how to design more robust GAs, and (3) shed more light on why mate selection matters in biology. The first four chapters of this thesis present motivations for this work, and describe investigations of the basic properties of mate selection in the context of GA. I employ the Schema Theorem and a Markov model as analytical tools to facilitate the study of mate selection. A number of empirical results are also presented to enhance our understanding of the GAs' behavior. The results based on simple test problems highlight the importance of mating choices in improving the GA's performance. Next, this study focuses on two classes of more complicated, building-block-based problems---the Royal Road functions and the hyperplane-defined functions. With the results further obtained, I introduce an important hypothesis regarding the role of mate selection in GAs. That is, if one's goal is to improve the GA's search for best-so-far solutions, then on easy problems a dissimilarity-based mate selection scheme is more beneficial. If problems present sufficient difficulty, the GA's search power can be further improved by reducing the selection pressure toward higher-fitness individuals while selecting mates. Chapter 6 presents a test of this hypothesis based on several more realistic, non-building-block-based benchmark testbeds. The test problems used are of increasing complexity in terms of various aspects of fitness landscapes. The first two testbeds, a sphere function and a step function, represent unimodal problems. The following four testbeds are multimodal in which characteristics of fitness landscapes such as deception and non-separability are included-the generalized Rosenbrock Saddle, an optimal control problem, a modified version of the Schaffer function F7 and Michalewicz's epistatic function. All the results of the experiments validate this hypothesis. This is encouraging-it implies that the ideas of mate selection proposed in this thesis can be applied to practical problems. Chapter 7 discusses a more general setting in the context of multimodal function optimization, engineering and machine learning. Identifying multiple peaks and maintaining subpopulations of the search space are two central themes. An immune system model is employed to study these two problems. The experimental results indeed shed more light on how mate selection schemes compare to traditional selection schemes. The final chapter provides a summary of this thesis, and highlights its contributions to GA research. It discusses paths for future research, and draws overall conclusions from the research presented in this thesis.Ph.D.Applied SciencesComputer scienceElectrical engineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/129644/2/3042087.pd