Development of Evolutionary Algorithms based on a Picture of Evolution of Probability Distribution

点击翻译按钮获取中文摘要

In this study, we aimed at Genetic Algorithms (GA) as optimization methods utilizing only function values to be optimized. We have examined the GA that uses population of search points with a picture of evolution of probability distribution, carried out comparison study of similar method called Estimation of Distribution Algorithms (EDA), and improved GA considering their applications to practical engineering problems.First, concerning comparative study between GA and EDA, we have proposed Pseudo-mutation and Pseudo-crossover as evaluation criteria for population-based probabilistic search algorithms. Then, using these criteria, we have evaluated GAs such as Simple GA, Spin Glass GA and Thermo-Dynamical GA and Bayesian Optimization Algorithm (BOA), a representative implementation of EDA.Further, from the viewpoint of evolution of distribution, we devised extension of real-coded GA for optimization of periodic function which is often appears in applications in engineering. It is based on the idea of embedding hyper sphere in the Euclidian space and applying the crossover in real-coded GA. Numerical experiments shows effectiveness of the proposed method.Since GA is applicable to optimization problems that involve noise, we have also applied GA to simulation-based optimization using random numbers. As a practical application, it has been applied optimization of group controller of elevator systems successfully. It can be also effective optimization tool for experiment-based optimization. In application of GA to elevator controller, implementation of controller requires decision making mechanisms and we have developed an exampler-based policy representation whose parameters are searched by GA. As well as simpler benchmarking problems, the exampler-based approach combined with GA works well in elevator control problem.