Research On Niche Genetic Algorithm Based On Fuzzy Cluster And Multi-Niche Crowding

Niche technology is widely used in genetic algorithm for maintaining the population diversity. Two improved niche genetic algorithms (niche genetic algorithm based on fuzzy similarity clustering and self-adaptive controlling of peaks radii, and multi-niche crowding genetic algorithm based on fitness sharing) are proposed in this paper, for overcoming the limitations of traditional multiple hump searching niche genetic algorithms. Several aspects are discussed in this paper as followed:1. GA's origins, its basic conceptions, general research circumstances and some foundation theories of GA, such as schema theorem, building block hypothesis, and implicit parallelism are mainly introduced.2. Crowding modal and fitness sharing modal, traditionally used in multi hump searching are analyzed thoroughly, and their limitations are proposed.3. Niche genetic algorithm based on fuzzy similarity clustering and self-adaptive controlling of peaks radii is proposed. The basic idea of the method is that, in the process of genetic evolvement, it takes the radii of peaks as a part of optimization variables, the radii of peaks are coded, put in the chromosomes and optimized with the variables of the problem by fitness sharing genetic algorithm without a prior knowledge of the above parameters, in the process of clustering, it controls the number of converged niches through adjusting the fuzzy similarity degree, avoiding finding the invalid extreme points as well. The algorithm takes no need to know the concrete number of niches and the value of the niche radium in advance, having a good searching ability on various multiple hump functions.4. Multi-niche crowding genetic algorithm based on fitness sharing is proposed. During the selection step, the algorithm uses the crowding selection policy, and during the replacing step, it uses a replacement policy called worst among most similar, after fitness sharing. For combines the idea of crowding and sharing, the algorithm has a good ability in escaping local optima, maintaining stable subpopulations in different niches, and converging to the global optima.5. Several classical test functions are carried on the two improved algorithms. Theoretical analysis and numerical experiments indicate that the algorithms keep a good diversity throughout the search, have a good searching ability in various multiple hump functions, and are less dependent as well.