The Research On Evolutionary Algorithm Ensembles For Global Numerical Optimization Problems

Abstract:In the field of scientific research and engineering application, a variety of problems should be done, and most of them can be attributed to the optimization problems. Evolutionary algorithms, which are widely used to solve optimization problems, are a class of random search methods, and show many excellent properties in the solving to practical problems and complex problems. However, according to the "no free lunch" theorem, there does not exist such an algorithm, which can solve all optimization problems effectively. In this case, many scholars began to devote to the research on evolutionary algorithm ensembles. Because that the evolutionary algorithm ensembles can show many excellent properties of their constituent algorithms, evolutionary algorithm ensembles can solve more problems than their constituent algorithms, that is to say evolutionary algorithm ensembles are a class of more universal methods.There are significant differences in the search mechanism among different evolutionary algorithms, so that there is natural heterogeneity between different algorithms, leading to different algorithms having different properties. Based on the analysis about the heterogeneity between two algorithms composite differential evolution (CoDE) and covariance matrix adaptation evolution strategy (CMA-ES), this thesis proposes a novel evolutionary algorithm ensemble named A Novel Evolutionary Algorithm Ensemble Based on CoDE and CMA-ES, which is abbreviated as EBCC.EBCC is a novel multi-method search algorithm, which uses CoDE and CMA-ES as constituent algorithms. In the progress of EBCC, the two constituent algorithms can learn from each other, in addition, according to the different characteristics of the two constituent algorithms, EBCC designs different methods to promote mutual learning for them. To more efficiently use the information of the past populations, EBCC saves the outstanding individuals of the two constituent algorithms in two different archives. Through maintaining the population diversity of the archives, there is a smaller probability for EBCC to stagnation and premature convergence. Meanwhile, EBCC can identify the performance of the two constituent algorithms dynamically, and EBCC will allot more function evaluations (FES) for the better one. In this case, EBCC can use the limited FES more effectively.The performance of EBCC has been tested on25benchmark test functions developed for the special session on real-parameter optimization of the2005IEEE Congress on Evolutionary Computation (IEEE CEC2005), and the result of the test show that EBCC is a more outstanding universal algorithm. Compared with other state-of-the-art evolutionary algorithms and the individual constituent algorithms, EBCC performs significantly better than all them, and all them show EBCC is excellent.