Rapid prediction of optimum population size in genetic programming using a novel genotype - Fitness correlation

The main aim of landscape analysis has been to quantify the 'hardness' of problems. Early steps have been made towards extending this into Genetic Programming. However, few attempts have been made to extend the use of landscape analysis into the prediction of ways to make a problem easy, through the optimal setting of control parameters. This paper introduces a new class of landscape metrics, which we call 'Genotype-Fitness Correlations'. An example of this family of metrics is applied to six real-world regression problems. It is demonstrated that genotype-fitness correlations may be used to estimate optimum population sizes for the six problems. We believe that this application of a landscape metric as guidance in the setting of control parameters is an important step towards the development of an adaptive algorithm that can respond to the perceived landscape in 'real-time', i.e. during the evolutionary search process itself. Copyright 2008 ACM.