Multiobjective data mining from solutions by evolutionary multiobjective optimization
Yusuke Nojima, Yuki Tanigaki, Hisao Ishibuchi · Proceedings of the Genetic and Evolutionary Computation Conference · 2017
One research direction in the field of evolutionary multiobjective optimization (EMO) is a post-analytical process of non-dominated solutions in order to analyze the relationship between design variables and objective functions for optimization problems. For this purpose, data mining techniques have been used in some studies. From a practical point of view, this process itself should be considered as a multiobjective optimization problem. In this paper, multiobjective genetic fuzzy rule selection is applied to the post-analytical process of solutions obtained by EMO algorithms. First, multiple regions of interest are specified in the objective space. Each region with a number of solutions is handled as a different class. A set of patterns is generated by the labeled solutions. Second, a number of fuzzy if-then rules are generated by classification rule mining. Finally, an EMO algorithm is applied to combinatorial optimization of fuzzy if-then rules in order to obtain a number of non-dominated fuzzy classifiers with respect to accuracy and complexity. Through computational experiments using two engineering problems, we show that we can obtain various classifiers with a variety of complexity-accuracy tradeoff.