Efficient search techniques using adaptive discretization of design variables on real-coded evolutionary computations
Toshiki Kondo, Tomoaki Tatsukawa · Proceedings of the Genetic and Evolutionary Computation Conference · 2018
In this paper, we evaluate the effects of adaptive discretization of design variables in real-coded evolutionary computations (RCECs). While the appropriate granularity of design variables can improve convergence in RCECs, it is difficult to decide the appropriate one in advance in most of the practical optimization problems. Besides, when the granularity is too coarse, the diversity may be lost. To address these difficulties, we propose two adaptive discretization techniques that discretize each design variable using granularity determined according to the indicator of solution distribution state in design space. In this study, standard deviation(SD) or estimated probability density function(ePDF) is used as an indicator for determining granularities of design variables. We use NSGA-II as an RCEC and thirteen benchmark problems including engineering problems. The generational distance (GD) and inverted generational distance (IGD) metrics are used for investigating the performance of convergence and diversity, respectively. To make sure the statistical difference of results, the Wilcoxon rank-sum test and Welch's t-test are applied in each problem. The results of experiments show that both of the proposed methods can automatically improve convergence in many problems. In addition, it is confirmed that the diversity is also maintained.