Adaptive weight vector assignment method for MOEA/D
Kei Harada, Satoru Hiwa, Tomoyuki Hiroyasu · 2017
MOEA/D is one of the multiobjective optimization methods in which an optimization problem is decomposed into subproblems and searches are performed. When the search difficulty of each subproblem is equal, the obtained Pareto solutions are evenly distributed. In contrast, in real problems, the search complexity of each subproblem is often not equal. In that case, a Pareto solution set to a difficult subproblem cannot be found by an easy search. To solve this problem, a method that adaptively assigns the weight vectors of MOEA/D according to the search situation is proposed. In the proposed method, subproblems that are difficult to search are divided into more subproblems, the search speeds of subproblems with different search difficulties are hence equalized, and solutions over a wider range should be found. The proposed method was used to a real-world problem, and its effectiveness is discussed. The target problem is to identify important brain regions using real data from a noninvasive functional brain imaging device. Identifying important brain regions is expected to promote elucidation of brain functions and contribute to effective training and therapeutic methods to improve human cognitive function. Compared to conventional MOEA/D, good solutions were obtained by the proposed method in the objective function space that is difficult to search. In addition, the influence of the proposed adaptive weight vector assignment was investigated, and it was confirmed that the proposed method adaptively allocates many weight vectors to the difficult search areas. Hence, the proposed method in this paper will extend the range of real problems that can be addressed using MOEA/D.