Visualizing Parameter Adaptation in Differential Evolution with Expected Fitness Improvement

Abstract

In this paper the expected fitness improvement metric is proposed to visualize the parameter search space in Differential Evolution. The expected fitness improvement is estimated at every generation of the algorithm and plotted in a heatmap profile. The spread of promising scaling factor values is analyzed for the SHADE and jDE algorithms with two different mutation strategies. In addition, the distance between the individuals in the population is considered, and the connection between distance and scaling factor values is observed. The performed experiments reveal important properties of Differential Evolution mutation operators, as well as widely used parameter adaptation techniques.