Revisiting Population Models in Differential Evolution on a Limited Budget of Evaluations

EasyChair Preprint 4218

Abstract

No previous study has reported that differential evolution (DE) is competitive with state-of-the-art black-box optimizers on a limited budget of evaluations (i.e., the expensive optimization scenario). This is true even for surrogate-assisted DE algorithms. The basic algorithmic framework of DE should be reconsidered to improve its performance substantially. In this context, this paper revisits population models in DE on a limited budget of evaluations. This paper analyzes the performance of DE with five population models on the BBOB function set. Results demonstrate that the traditional synchronous model is unsuitable for DE in most cases. In contrast, the performance of DE can be significantly improved by using the plus-selection model and the worst improvement model. Results also demonstrate that DE with a suitable population model is competitive with covariance matrix adaptation evolution strategy depending on the number of evaluations and the dimensionality of a problem.

Keyphrases: Computationally expensive optimization, Differential Evolution, Population models

@booklet{EasyChair:4218,
  author    = {Ryoji Tanabe},
  title     = {Revisiting Population Models in Differential Evolution on a Limited Budget of Evaluations},
  howpublished = {EasyChair Preprint 4218},
  year      = {EasyChair, 2020}}