@techreport{oai:ipsj.ixsq.nii.ac.jp:00218934,
 author = {Jiayi, Han and Shinya, Watanabe and Jiayi, Han and Shinya, Watanabe},
 issue = {7},
 month = {Jul},
 note = {Multi-objective evolutionary algorithm based on decomposition (MOEA/D) is a powerful algorithm and provides a framework for solving multi-objective optimization problems (MOPs). Differential evolution (DE) algorithm and its variants are often used to generate new solutions in MOEA/D heuristically. However, based on the “No Free Lunch” theory, only a fixed algorithm for generating new solutions in the original MOEA/D cannot efficiently solve all MOPs. Therefore, in this paper, we propose a new framework based on MOEA/D named MOEA/D-EF (Ensemble Framework), which can contain a variety of new-solutions generating algorithms (candidate algorithms) with different search capabilities to improve the overall universality of the algorithm. In the new approach, the whole iteration is divided into the evaluation generation (EG) and the implementation generation (IG). We provide a fair evaluation environment for each candidate algorithm at the beginning of each generation belonging to the EG and evaluate their performance by using the Hypervolume indicator. The algorithm with the best performance in one EG will be chosen and executed in the following IG. Also, we believe that some historical information representing evolutionary details can help generate superior new solutions. Thus, in numerical experiments, we take our original DE variant based on the ideal point and historical information as one of the candidate algorithms for generating new solutions. The numerical experiments show that the new framework has broader universality., Multi-objective evolutionary algorithm based on decomposition (MOEA/D) is a powerful algorithm and provides a framework for solving multi-objective optimization problems (MOPs). Differential evolution (DE) algorithm and its variants are often used to generate new solutions in MOEA/D heuristically. However, based on the “No Free Lunch” theory, only a fixed algorithm for generating new solutions in the original MOEA/D cannot efficiently solve all MOPs. Therefore, in this paper, we propose a new framework based on MOEA/D named MOEA/D-EF (Ensemble Framework), which can contain a variety of new-solutions generating algorithms (candidate algorithms) with different search capabilities to improve the overall universality of the algorithm. In the new approach, the whole iteration is divided into the evaluation generation (EG) and the implementation generation (IG). We provide a fair evaluation environment for each candidate algorithm at the beginning of each generation belonging to the EG and evaluate their performance by using the Hypervolume indicator. The algorithm with the best performance in one EG will be chosen and executed in the following IG. Also, we believe that some historical information representing evolutionary details can help generate superior new solutions. Thus, in numerical experiments, we take our original DE variant based on the ideal point and historical information as one of the candidate algorithms for generating new solutions. The numerical experiments show that the new framework has broader universality.},
 title = {A new Ensemble Framework based on MOEA/D},
 year = {2022}
}