Operator and parameter adaptation in genetic algorithms

Abstract

 Genetic Algorithms are a class of powerful, robust search techniques based on genetic inheritance and the Darwinian metaphor of “Natural Selection”. These algorithms maintain a finite memory of individual points on the search landscape known as the “population”. Members of the population are usually represented as strings written over some fixed alphabet, each of which has a scalar value attached to it reflecting its quality or “fitness”. The search may be seen as the iterative application of a number of operators, such as selection, recombination and mutation, to the population with the aim of producing progressively fitter individuals.

These operators are usually static, that is to say that their mechanisms, parameters, and probability of application are fixed at the beginning and constant throughout the run of the algorithm. However, there is an increasing body of evidence that not only is there no single choice of operators which is optimal for all problems, but that in fact the optimal choice of operators for a given problem will be time-variant i.e. it will depend on such factors as the degree of convergence of the population. Based on theoretical and practical approaches, a number of authors have proposed methods of adaptively controlling one or more of the operators, usually invoking some kind of “meta-learning” algorithm, in order to try and improve the performance of the Genetic Algorithm as a function optimiser.

In this paper we describe the background to these approaches, and suggest a framework for their classification, based on the learning strategy used to control them, and what facets of the algorithm are susceptible to adaptation. We then review a number of significant pieces of work within the context of this setting, and draw some conclusions about the relative merits of various approaches and promising directions for future work.