A competitive mechanism based multi-objective differential evolution algorithm and its application in feature selection

A large number of evolutionary algorithms have been introduced for multi-objective optimization problems in the past two decades. However, the compromise of convergence and diversity of the nondominated solutions is still the main difficult problem faced by optimization algorithms. To handle this problem, an efficient competitive mechanism based multi-objective differential evolution algorithm (CMODE) is designed in this work. In CMODE, the rank based on the non-dominated sorting and crowding distance is first adopted to create the leader set, which is utilized to lead the evolution of the differential evolution (DE) algorithm. Then, a competitive mechanism using the shift-based density estimation (SDE) strategy is employed to design a new mutation operation for producing offspring, where the SDE strategy is beneficial to balance convergence and diversity. Meanwhile, two variants of the CMODE using the angle competitive mechanism and the Euclidean distance competitive mechanism are proposed. The experimental results on three test suites show that the proposed CMODE performs better than six state-of-the-art multi-objective optimization algorithms on most of the twenty benchmark functions in terms of hypervolume and inverted generation distance. Furthermore, the proposed CMODE is applied to the feature selection problem. The comparison results on feature selection also demonstrate the efficiency of our proposed CMODE. (c) 2022 Elsevier B.V. All rights reserved.

Automated summary

This paper proposes an efficient competitive mechanism based multi-objective differential evolution algorithm (CMODE) to address the compromise between convergence and diversity in optimization algorithms. The CMODE utilizes rank-based non-dominated sorting and crowding distance to create a leader set, and a competitive mechanism using shift-based density estimation is employed to balance convergence and diversity. Experimental results demonstrate that the proposed CMODE outperforms other algorithms on benchmark functions and feature selection problems.