Hybrid genetic algorithms with reduced premature convergence for search performance enhancement

Abstract

This thesis deals with the search performance enhancement techniques in genetic algorithms (GAs). Evolutionary algorithms are computational optimization techniques that use simulated evolution. In control area, especially the genetic algorithm has been widely used. Although GAs are good at finding near global optimum quickly, they are poor in the fine tuning of solutions, which may cause `premature convergence``. This thesis is aimed to improve the search performance of GAs with elitist strategy by reducing premature convergence through hybridization or appropriate modification to the algorithms. In order to accomplish the goals, three new genetic algorithms and two new local search operators are proposed. The first method, the modified genetic algorithm (MGA) consists of a fitness modification scheme and adaptive mutation operator. The second method, the adaptive genetic algorithm (AGA) determines crossover and mutation probabilities by itself according to the fitness of a solution to be crossed or mutated. The schema theorem for AGA is derived. The third method, adaptive simulated annealing genetic algorithm (ASAGA) uses simulated annealing-like mutation operator. A novel way of generating a new solution by using a gaussian random number with time-varying variance is proposed and proved to be effective. The first local search operator makes use of neural networks. The second one named SLSO (simple local search operator) uses the difference between the most recent best fitness and a newly found best solution``s fitness, which is computationally simpler than the first one and proved to be powerful. The test problems considered for the performance comparison include the traditional set of test functions, system identification, neural network controller for cart-pole system, evolutionary design of a multi-agents playing a simplified soccer and nonlinear constrained optimization problems. Nine hybrid genetic algorithms are constructed using the proposed algorithms...