Predictive Models of Problem Difficulties for Differential Evolutionary Algorithm Based on Fitness Landscape Analysis

Abstract

The differential evolutionary (DE) algorithm is one of the evolutionary algorithms widely used to solve various optimization problems. The problem of how to set the parameters of DE has not still been solved effectively. A large number of solutions need to be generated during the running of the algorithm before signs of convergence appear. In this paper, we focus on the relationship between specific algorithm-independent problem characteristics and the behaviour of the algorithm in response to specific parameter setting. We use fitness landscape analysis to identify and extract those problem difficulties' features which can also reflect the relationship between the chosen parameters of DE and the generated solutions. We use decision tree induction to design predictive performance models and demonstrate the success of these models in indicating the success or otherwise of a specific parameter setting. Linear regression is also employed to analyse the experimental results.