The Research On Noisy Evolutionary Multi-Objective Optimization Algorithm Based On Mixed Selection

Multi-objective optimization has been successfully used in many fields.Evolutionary algorithm as a kind of heuristic searching algorithm is often used to deal with multi-objective optimization problems by researchers.Multi-objective optimization problems usually involve noise in real world.The noise may appear in random model parameters,objective functions and decision variables with different distributions,which has a certain impact on the multi-objective optimization problems to be studied.When we evaluate noisy fitness function,the attaining objective value will become larger or smaller compare to real objective values.Noise makes the function value fraudulent which let individuals entering the next generation can’t be precisely selected.This thesis proposes three kinds of structure of noisy multi-objective algorithms based on mixed selection mechanism.The three algorithms are developed based on two selection ways that one is non-dominating ranking selection,another is probability ranking selection.The detailed work arrangement of this paper is as follows:1.A noisy evolutionary multi-objective optimization algorithm with regular model based on mixed selection is proposed.The solutions establishing the regular models in this algorithm are from the union set of solutions selected by non-dominated sorting and probabilistic sorting.In this paper,the number of solutions chosen by the two methods is respectively determined by the number of crossing solutions from which method in the contemporary population and in the first generation,we need to initialize this two numbers.Finally,sampling population attaining from regular models join in the process of evolutionary.To verify the effectiveness of the algorithm,we compare our algorithm with other four algorithms and the results shows that our algorithm has certain advantages in dealing with multi-objective optimization problems in noisy environment.2.A noisy evolutionary multi-objective optimization algorithm based on improved regular model is proposed.The difference between this algorithm and the first work is that only the solutions of probability ranking is used to build the regular model and obtain the sampling solutions in this algorithm and the solutions of non-dominating method is not used for modeling,but directly combines with the sampling solutions to form a new population.In this way,the solutions obtained by the non-dominating method can improve the diversity of the algorithm in the evolutionary process,and help the evolutionary direction of the algorithm as close as possible to the optimal.The experimental results of this algorithm also prove that the algorithm has much better the ability of denoising noise than the first one,especially for some multimodal problems.3.A noisy evolutionary multi-objective optimization algorithm with regular model based on mixed two kinds of models is proposed.The difference between the algorithm of this work and the algorithm of the second work is that the obtaining solutions by the non-dominating way are used to build gaussian models and attain sampling solutions on gaussian models,but the solutions obtaining by probabilistic ranking is still used to build regular models and attain sampling solutions on the regular models.The sampling solutions of these two parts are combined to form a new population.Compared with the first two algorithms,this algorithm has certain advantages.