Using Multivariate Quantitative Genetics Theory to Assist in EA Customization

Customizing and evolutionary algorithm (EA) for a new or unusual problem can seem relatively simple as long as one can devise an appropriate representation and reproductive operators to modify it. Unfortunately getting a customized EA to produce high quality results in a reasonable amount of time can be quite challenging. There is little guidance available to help practitioners deal with this issue. Most evolutionary computation (EC) theory is only applicable to specific representations, or assumes knowledge of the fitness function, such as the location of optima. We are developing an approach based on theory from the biology community to address this problem. Multivariate quantitative genetics theory characterizes evolving populations as multivariate probability distributions of phenotypic traits. Some advantages it offers are a degree of independence from the underlying representation, and useful concepts such as phenotypic heritability. Re-working the quantitative genetics equations, we expose an additional term that we call "perturbation". We believe that perturbation and heritability provide quantitative measures of the exploration and exploitation, and that practitioners can use these to identify and diagnose imbalances in customized reproductive operators. To illustrate, we use these tools to diagnose problems with a standard recombination operator for a Pittsburgh approach classifier system. With this knowledge we develop a new, more balanced, recombination operator, and show that its use leads to significantly better results.