Expensive Multi-Objective Evolutionary Algorithm with Multi-Objective Data Generation

Expensive multi-objective optimization problem is a class of complex optimization problems that involve multiple objectives that are conflicted with each other and computationally expensive, which requires algorithms to find a set of non-dominated solutions as many and diversity as possible with limited computational resources. Although evolutionary computation algorithms are regarded as effective tools for solving multi-objective optimization problems, they still face the diversity and convergence challenges when solving expensive multi-objective optimization problems, i. e., difficulty to find a set of solutions that are of good diversity and converged to Pareto front. To address the diversity and convergence challenges, this paper proposes a novel multi-objective data generation-based expensive multi-objective evolutionary algorithm. The contributions and innovations of this paper can be mainly summarized in the following three aspects. First, this paper puts forward and proves the non-dominated solution generation theorem, and then proposes a multi-objective data generation method based on the theorem, so as to obtain more non-dominated solutions more efficiently for improving the solution diversity. Second, this paper proposes a multiple population for multiple surrogates framework that co-evolves multiple populations to efficiently optimize multiple surrogates that are built for multiple real expensive objectives respectively. Third, based on the above proposed method and framework, this paper proposes the expensive multi-objective evolutionary algorithm based on multi-objective data generation for efficiently solving the expensive multi-objective optimization problem. To validate the algorithm performance, extensive experimental analyses are conducted on 16 problems from two well-known test sets in the related field with five existing state-of-the-art algorithms as competitors in this paper. The experimental results show that the algorithm proposed in this paper is able to achieve better metric values than all the compared algorithms on most of the problems, with good effectiveness and efficiency. © 2023 Science Press. All rights reserved.