An Improved Multi-Population Genetic Algorithm for Constrained Nonlinear Optimization

Penalty function is popular method for constrained optimization problems. Generally, a penalty parameter controls the degree of penalty for a constrained violation and an optimal parameter exists, but the value is difficult to define and its optimal value is different for different questions. Here, we propose an improved multi-population genetic algorithm to solve this problem. Each population uses different penalty strategy, then each subpopulation evolve independently for a certain number of generations, after that exchange individuals between different subpopulations. This method can perform multi-directional searches by manipulating several subpopulations of potential solutions for different penalty degree for constraints violation and obtain mixed information from these different directional searches, so it can make the selection of the penalty degree much easier and has more chance to find an optimal solution.