An improved regularity-based vector evolutionary algorithm for multi-objective optimizations. (February 2023)

• Record Type: Journal Article
• Authors: Li, Yiying; Yang, Shiyou
• Is Part Of: Advances in engineering software. Volume 176(2023)
• Publication Date: 2023-02
• Subjects: Evolutionary algorithm -- Multi-objective optimization problem -- Interpolation -- Inverse models; Computer-aided engineering -- Periodicals; Engineering -- Computer programs -- Periodicals; Engineering -- Software -- Periodicals; Periodicals; 620.0028553
• DOI: 10.1016/j.advengsoft.2022.103397
• Languages: English
• ISSNs: 0965-9978

Abstract

Highlights: Inverse models-based evolutionary algorithms can generate as many solutions as required in interest regions at a low budget cost. The number of predefined reference vectors remains constant resulting in a low computational efficiency. The non-dominated solutions are utilized for interpolations in each subpopulation. An adaptive interpolation range coefficient for adaptive adjustments of the interpolation region is to balance the exploitation and the exploration. Hiring elite individuals in the whole population to enhance the solution convergence. Abstract: Inverse models-based evolutionary algorithms can generate as many solutions as required in interest regions at a low budget cost. However, the objective value used for the inverse model in the original regularity-based algorithm using Gaussian process-based inverse models, (IM-MOEA), is linearly and separately sampled for each objective according to its maximum and minimum values so far searched. Consequently, some non-optimal even infeasible solutions may be sampled. Also, the number of the predefined reference vectors remains constant throughout the whole evolution. Consequently, it may result in a low computational efficiency for multi-objective optimization ones (MOOs) with nonuniform or disconnected Pareto fronts. In this respect, an improved regularity-based vector evolutionary algorithm for multi-objective optimizations is presented in the paper. The main component of the proposed algorithm is to obtain the training data for inverse models by interpolation in the objective space. The population is divided into subpopulations according to the number of objectives. The non-dominated solutions are utilized for interpolations in each subpopulation. The inverse models are finally used to generate offspring using data by interpolation. An adaptive interpolation range coefficient is proposed for adaptive adjustments of the interpolation region to balance the exploitation and the exploration searches. An elite strategy of hiring elite individuals in the whole population to the current subpopulation is employed to enhance the algorithm convergence. Experimental results on two test suites show the superiority of the proposed algorithm.