BS-AWA: Adaptive Weighted Aggregationの目的数に対するスケーラビリティの向上

This paper proposes a more scalable variant of Adaptive Weighted Aggregation (AWA) with respect to the number of objectives in continuous multiobjective optimization. AWA is a scalarization-based multi-start strategy for generating finite points that approximate the entire Pareto set and Pareto front, which is especially focused on many-objective problems (having four or more objectives). In our last study, we discussed a reasonable stopping criterion for AWA, the representing iteration, and analyzed the time and space complexity of AWA when the representing iteration is used as a stopping criterion. Theoretical and empirical results showed that the running time and memory consumption of AWA depends on the number of solutions found in the representing iteration, the representing number. Due to the factorial increase of the representing number for objectives, the applicability of AWA is limited to 16-objective problems. In this study, we therefore redesign two central operations in AWA, the subdivision and the relocation, in order to reduce the representing number. The new subdivision is based on the simplicial complex and its barycentric subdivision and the new relocation is based on the simplicial approximation of a mapping and its range, both of which are well-known notions in topology. We theoretically compare the new AWA, named the barycentric subdivision-based AWA (BS-AWA), with the old AWA in terms of their representing iteration, representing number and approximate memory consumption to illustrate the improvement of scalability; the result implies that BS-AWA is applicable to over 20-objective problems. Numerical experiments using 2- to 17-objective benchmark problems show that BS-AWA achieves a better coverage of obtained solutions than conventional multi-start descent methods in both the variable and objective spaces. The running time and the solution distribution of BS-AWA are also discussed.