2010 年 1 巻 1 号 p. 65-78
This work proposes a hybrid strategy in a two-stage search process for many-objective optimization. The first stage of the search is directed by a scalarization function and the second one by Pareto selection enhanced with Adaptive ε-Ranking. The scalarization strategy drives the population towards central regions of objective space, aiming to find solutions with good convergence properties to seed the second stage of the search. Adaptive ε-Ranking balances the search effort towards the different regions of objective space to find solutions with good convergence, spread, and distribution properties. We test the proposed hybrid strategy on MNK-Landscapes and DTLZ problems, showing that performance can improve significantly. Also, we compare the effectiveness of applying either Adaptive ε-Ranking or NSGA-II's non-domination sorting & crowding distance in the second stage, clarifying the necessity of Adaptive ε-Ranking. In addition, we include a comparison with two substitute assignment distance methods known to be very effective to improve convergence on many-objective problems, showing that the proposed hybrid approach can find solutions with similar or better convergence properties on highly complex problems, while achieving better spread and distribution.