Some Improvements And Applications Of Evolution Strategies

In order to solve the problems of evolution strategies such as premature convergence, falling into local optima and low accuracy, in recent years, many experts and scholars have carried out some researches and presented a lot of soluntions. However, those problems still hasn't been dealed with successfully. Furthermore, some problems also occurred in multimodal optimization and large-scale optimization. Our research is focused on this field. In this paper, those problems are analyzed, and by combining some main methods of machine learning we present an improved algorithm. The improved algorithm effectively combines the respective advantages and can effectively solve some of the problems that existing in evolution strategy. This paper presents a Gaussian classifier-based evolutionary strategy (GCES) to solve the multimodal optimization problems. It can locate local or global optima and provides better and consistent performance; In order to solve the large-scale optimization problems, this paper combines several evolutionary strategy models by cooperative co-evolution, from a certain extent, it can accelerate the search speed and improve the global optimization performance; at the same time, the paper applied the evolutionary strategy to predict the evolution of the underwater terrain of the riverbed, and got satisfactory prediction effect. The main work is summarized as follows:First, in order to locate the multiple optima of multimodal optimization problems, this paper proposes a new MOPs strategy based on the Gauss classifier, named the gaussian classifier based on the evolutionary strategy (GCES). We studied several methods to estimate the covariance of GMM, and and adjust a zoom factor (ZF) that is equivalent to the global step size in the existing evolution strategy (ES) family, to balance between exploration and exploitation. The idea behind GCES is that each individual, i.e., each feasible solution in evolutionary population, is first classified into different clusters according to the Bayesian posterior probability based on the GMMs of currently found clusters. Then, every GMM will be updated independently at every generation based on the results of the classification. To simplify the calculation, the prior probability of each GMM is assumed to be the same. This process continues until meet the stop criteria. This method is similar to the niche radius adaptive with the covariance matrix adaptive ES (CMA-ES). However there are many differences with the most marked one being the distinction in the concept. In GCES, the Gauss classifier is introduced as a niche instead of the niche radius, and in the CMA-ES, the niche radius and distance among individuals is used. Without the radius constraint, the proposed method concentrates on only the classifier model and estimation strategy of the probability distribution, thereby requiring much fewer parameters to be set than CMA-ES and its variants.Secondly, a large number of practical engineering optimization problems are usually high dimensionality ones, and the current evolutionary algorithms are mostly designed for the below 1000 dimensional optimization problems, so the research of the high dimensional optimization problem has practical significance. We have adopted a hybrid evolutionary strategy algorithm based on the collaborative mechanism to solve the large-scale optimization problems:first of all, the optimization problem is decomposed; then, several evolutionary strategy algorithms to solve the problem together. The divide and rule strategy is a good ideal for solving large-scale optimization problems. By studying the recent popular decomposition strategy called Differential Grouping, we improved it and proposed our solution: Disjoint Differential Grouping. Then we proposed a hybrid evolutionary strategy named Cooperatively Coevolving Evolutionary Strategies which includes (?,?) and (?+?) to cooperate with each other to solve the high dimensional optimization problem.Finally, the paper researched the prediction problem of the underwater terrain of the riverbed. This problem requires the use of historical data, analysis of the evolution of the middle reaches of the river, summed up the relative regularities. And by using the theories, means and methods of recent science and technology, it requires simulating the changes of the shoal, predicting the future shoal in a certain period of time and providing accurate guidance for the channel maintenance and regulation.That is particularly important at the present stage. First of all, we extract the feature of the data, build the model by using time series and cellular automaton methods, then solve the parametric model by the methods of evolution strategy, simulate the riverbed evolution process and predict the evolution trend of the riverbed, provide scientific guidance for waterway maintenance. So the research of this project which can guide the waterway maintenance of the Yangtze River is very important and has far-reaching significance either in theory or in practice.Some suggestions that proposed for improvement could greatly improve the performance of the evolutionary strategy and expand the application fields of evolution strategy. Simulation experiment results show that these improvements are feasible, with the characteristics of fast convergence speed, robustness and stability.