Selective crossover as an adaptive strategy for genetic algorithms

Since the proposal of the first genetic algorithm (GA) many recombination operators have been proposed. Some are problem specific and require a great deal of knowledge about the problem being solved, resulting in good but highly specialised operators. Other recombination operators have been proposed for more general use. One advantage for such operators is the little knowledge required about the problem being solved; however, the synergy of these operators, the problem being solved and other GA parameters does not always yield optimum performance from the GA. More recently, adaptive recombination operators have been proposed to bridge the gap between general and specialised recombination operators. This thesis presents a novel adaptive recombination operator, namely "Selective Crossover", for use with a genetic algorithm. Selective crossover was designed with three properties that make it a viable strategy to use when little or no knowledge is available about the problem being optimised. The first property is the identification of allele changes made to the candidate solution during recombination. The second property is the use of correlations between parental and offspring fitnesses to discover beneficial alleles. The third property is the preservation of alleles at each locus, during recombination, according to their previous contributions to beneficial changes in fitness. This thesis makes six contributions. The first is the design and implementation of selective crossover. The second is a measurement and comparison of the performance of selective crossover and two traditional recombination operators on a number of different problems. The third is an empirical analysis of the adaptive properties in selective crossover. The fourth is an identification and analysis of four key biases inherent in selective crossover and a demonstration of the existence of these biases in two other similar operators. The fifth is an analysis and comparison of schema propagation in selective crossover and two traditional recombination operators. The final contribution is a construction of a schema survival probabihty for selective crossover.