Research On Many-objective Evolutionary Algorithm Based On Surrogate And Clustering Selection

The objective space dimension of many-objective optimization problems is higher,and the proportion of non-dominanted solutions increases,resulting in lower selection pressure among populations with limited number of individuals.It is difficult to effectively select solutions by traditional Pareto dominance-based method.In order to solve many-objective optimization problems,the utilization of the many-objective evolutionary algorithms is an effective method.The objective functions of many real problems are very complex,and the calculation of the function values is very time consuming.Therefore,the objective function can be replaced approximately using surrogate model.This paper proposes a surrogate model and clustering selection based evolutionary algorithm(SCSEA)for expensive many-objective optimization problems.The proposed algorithm is based on an improved Two-Archive algorithm for many-objective optimization that is named Two-Archive2.In this paper,various conditional characteristics are designed by analyzing and utilizing factors such as model errors and relative errors.An adaptive evolutionary control strategy is proposed based on decision tree,and the evaluation quality of the model is used to determine whether the real fitness function will be used.The Kriging models are used to approximate each objective function for obtaining sub-optimal solutions with reduced computation cost.The selection of training data in model management directly affects the quality of the model.This paper proposes to use clustering method for choosing representative individuals as samples to train models.In addition,the update mechanism of surrogate model is designed.We need to control the size of the training samples in order to avoid the high computational complexity.As too many or too few training data will affect the performance of the algorithm,different from the traditional methods of using a fixed number of training data,we propose a dynamic sample management method.This paper applies the proposed SCSEA algorithm to a series of benchmark functions with different objective numbers,additionally,empirical study on comparing the novel algorithm with five other evolutionary algorithms is conducted.The experimental results show that the proposed algorithm performs significantly better than the other evolutionary algorithms.