Research On Adaptive Evolution And Compound Crossover's Application In Genetic Algorithm

Genetic Algorithm is a general, robust, randomized heuristicsearch technique which is widely applied to auto-control,combinatorial optimization, image processing, robots, artificial life,machine learning, artificial intelligence, engineering design and so on.Especially when problem search space is very large, complicated orwe know little about problem in advance, it is inappropriate forconventional search tools (such as enumeration method and heuristictechnique) to solve these problems, but Genetic Algorithm can doeffectively and efficiently. At first, this paper introduces recent research works and resultsabout improving GA's search performances up to now. In this paper,we propose a similarity-based genetic algorithm (SBGA). It modifiescrossover and mutation strategy of Simple Genetic Algorithm (SGA).When parents' similarity degree (SD) is lower, crossover operator isused to generate children, otherwise mutation operator is used. On theone hand, superior individuals generated by crossover operatorneedn't go through mutation and can survive after competitions withother individuals generated by crossover and mutation operators. Butfor simple Genetic Algorithms, superior individuals generated bycrossover operator have to undergo mutation and may turn to beinferior individuals after mutation procedure, then influence GA'sconvergence quality and convergence speed. On the other hand,SBGA doesn't allow higher similarity individuals to crossover, soprevent from the loss of population's diversity by crossover andrestrain premature convergence in a certain extent. Then this modifiedgenetic algorithm is applied to figuring out several benchmark testfunctions, and we find that it outperforms genetic algorithm whichdoesn't employ similarity information. We use the algorithm toconstruct degree-constrained minimum spanning tree (d-MST) of agraph. Our experimental results provide strong evidence that thegenetic algorithm employing similarity information to decidecrossover or mutation finds significantly lower-cost solutions torandom graph d-MST problems than rival method. At last, wedescribe an all-purpose compound crossover which operates on tworeal-coded decision variables. When applied to higher-dimensionmulti-objective optimization problems, it performs very well.