A Study On The Multi-Population Evolutionary Algorithms For Discrete Variables

The genetic algorithm (GA) and cross entropy method (CE) are both effective technologies with global random search. However, they are prone to slow or premature convergence with a single population optimization. In this dissertation, the dual-population model in the team progress algorithm (TPA) is introduced for GA and CE improvements. Based on the introduced principle, the dual-population genetic algorithm and the dual population cross entropy method are presented respectively.The mutation of GA and the whole operation process of CE are both similar to the exploration action of TPA. So the two dual population algorithms can be designed if only the study action and member renewal rules of TPA are introduced. The test results for benchmark functions show that the convergence speed and global optimization capability of the two new algorithms are significantly improved. For the TSP, the DPCE using CE model of the combinatorial optimizations and the DPGA using 2-opt neighborhood search strategy are both transformed to the discrete variable optimization algorithms, respectively. After modification they can give the optimal tours for the TSP less than 30 cities. And the best solutions obtained for the multi-city problems have better qualities than the single algorithms. At last, the presented algorithms are used for side lobe level (SLL) optimization of phased arrays. The element excitation phases are continuously and discretely optimized, respectively. By optimizing the 20 element liner array and 64 element planar array under different steering angles, lower SLL can be obtained by use of the discrete optimization.With the behavior division of the two groups, the presented algorithms are both capable of fast convergence and global optimization. The results for function optimization and discrete applications verify that both the new algorithms also have the properties of strong universal, low computational cost, high stability and easy tuning of parameters, and can solve a variety of practical problems. Therefore, the proposed dual-population scheme in this dissertation will become a new approach for algorithm improvement.