Niching genetic algorithms exploring structure of landscape

This paper proposes a kind of parallel local search operator PLS and a restricted crossover operator GC having self_learning ability of the structure of fitness landscape, and qualitatively analyzes the operation mechanisms. Several formulas based on empirical data for calculating convergence velocity and global convergence reliability are firstly presented. A large number of experiments on several typical functions show that PLS is not only able to sharply speed up searching but also sufficiently maintain genotypic diversity in populations, and GC can precisely tune solutions in the late stage of searching. Both the convergence velocity and the global convergence reliability of the improved GAs introducing PLS and GC excel greatly that of standard ones, and have good robustness and stability.