The Study On Mechanism And Improvement Of Real-coded Genetic Algorithms

Genetic algorithms, usually adoptting real-coded, as one of the important branch of evolutionary algorithms, have been used to solve complicated optimization problems more and more widely. However the theoretical results on mechanism research of real-code genetic algorithms (RCGA) was relatively few, which was studied and several improved GAs were proposed in this dissertation. The main contributions are as followes:1,Action mechanism of genetic operators was analysed.There was negative correlation between fitnees value and the distance between global optimal solution and individuals. One pass of simple cross operator is a one-dimensional search based on differential method in essence. One generation of simple cross operation can be regarded as a combination of effective neighborhood search and local random search. The proportion of neighborhood search will change with the number of advantage block and the size of advantage block in population. New individuals generated by single- (or multi-) uniform (or nonuniform) mutation operator occupied problem space with non-uniform distribution. Enhancing seeds number of neighborhood of advantage individuals will improve the local convergence rate.2,A new analytical method of genetic drift of genetic algorithm with real-code was proposed by studying the expectation size growth of advantage subpopulation. The effect of genetic operator on the expectation size of advantage subpopulation was estimated concerned local characteristics of problems. Block Theory of RCGA was deduced that the processing of RCGA is expressed. After random searching of the first phase, the size of advantage blocks increase at the exponent growth but only optimal subpopulation was survived with the creasing of average fitness value and whose size tended to a steady value. The difference between Block Theory and Schema Theory was contrasted. The appropriate parameter setting was analyzed and the prematurity reason was explained. From the macroscopic point of view, RCGA is a iteration method to select multiple blocks at a diverse probability to the next generation.3,Based on the analysis of advanced convergence by block theory, The dissertation proposed a novel genetic algorithm with several elitists preserved method(GAEP). After analysing on the limitations of GAEP, Genetic algorithms with speices selection (GASS) come into being. Improved GASS(IGASS) was proposed by using stage characteristic of biological evolution for reference. Local searching was enhanced by boundary auto-contraction of optimal subpopulation. Gene segregation was carried out by the size of optimal subpopulation decreasing aperiodically and cross operation between optimal subpopulation and rest subpopulation. Optimization problems with separable variables were defined whose properties were studied. It's analyzed that the three algorithms are effective and especially appropriate for solving high-dimension optimization problems whose variables are separable, which is approved by numerical experiments.4,Common constraint-handling techniques are surveyed, which were not combined with elitist preservation strategies. A new constraint-handling techniques was proposed combined elitist preservation strategies and penalty function method, whose penalty factor was designed by local analysis. A new memetic algorithms to slove constraint optimization problems come into being from IGASS. Some numerical tests were made. The results show that the algorithm is effective and robust.