A coevolutionary estimation of distribution algorithm based on dynamic differential grouping for mixed-variable optimization problems

Abstract

The mixed variable optimization problems (MVOPs), which involves both continuous and discrete decision variables, are difficult to be solved due to the complex search space. Recently, many EA-based algorithms have been designed to address MVOPs. However, due to the mixed variables with different evolutionary operators and complex search space, it is difficult to handle the mixed variables effectively and the search efficiency cannot be guaranteed. How to solve MVOPs efficiently has been a challengeable issue. In this paper, we propose a mixed-variable optimization algorithm called coevolutionary estimation of distribution algorithm (CoEDAmv). First, a dynamic differential grouping (DDG) method is employed to improve the search efficiency of CoEDAmv, in which both the interaction of variables and search performance on the current search region are considered simultaneously. Second, two probabilistic models, i.e. fitness rank based continuous histogram (FRCH) and fitness rank based discrete histogram (FRDH), are proposed to handle continuous and discrete variables respectively, which can benefit from elite individuals obtained during the fitness ranking strategy and enhance the convergence performance with the elite neighborhood-based updating probability strategy. Compared with eight state-of-the-art algorithms, the experimental results on 28 artificial MVOPs show that CoEDAmv is effective and efficient.