Research On Dynamic Multi-objective Optimization Algorithm Based On Learning

Dynamic multi-objective optimization problems(DMOPs)widely exist in industrial production and real life.As the problem to be optimized in DMOPs changes continuously with time or environment,tracking the changing Pareto sets or Pareto front rapidly and efficiently is the key to solve DMOPs.Due to their independence from the nature of the problem as one of the fastest and most efficient algorithms,Dynamic multi-objective evolutionary algorithms(DMOEAs)are widely used to solve DMOPs.The design of response strategies to solve problems that change with the environment or time in a timely and efficient manner have attracted the attention of many researchers in recent years.In this thesis,two DMOEAs are proposed based on the advantages and limitations of the DMOEAs.The main studies of this thesis are as follows:1.Most existed DMOEAs focus on learning environmental change patterns from correlations between solution sets in the historical environment to build predictive models.However,information embedded within the solution sets in the historical environment such as the neighbourhood distribution of individuals and their local correlations in the decision space are ignored,which may affect the accuracy of prediction and the quality of the predicted population.Therefore,a manifold clustering-based optimization algorithm is proposed in this thesis.The proposed method firstly finds the internal linear neighborhood relations of individuals,then local linear manifolds are extracted from historical solutions.Individuals in the historical population will be grouped into several clusters based on the different linear manifolds to which the individuals are attached.Individuals belonging to a cluster are linearly related and have similar movement trends.Finally,the initial population in the new environment is composed of each initial subpopulation predicted by multiple prediction models established based on each cluster.The proposed manifold learning-based prediction is tested on some commonly used benchmark problems and compared with some state-of-the-art algorithms.The results confirm the effectiveness of the proposed method,especially in solving DMOPs with non-linear correlation between decision variables.2.The effectiveness of the proposed manifold clustering-based optimization algorithm may not be superior in problems with linear correlations,possibly due to the absence of guidance from the information of the new environment.This may lead to a lack of accuracy in the predicted direction.Furthermore,the proposed algorithm is population-based and pay more attention on extracting and utilizing the relationships between final populations in historical environments,while overlooking the correlation among individuals within the same environment.To address these problems,improve the performance of DMOPs by extracting and utilizing historical information,an inter-individual correlation and dimension-based dual learning method is proposed in this thesis.Two learning strategies,decomposition-based interindividual correlation transfer learning and dimension-wise learning,are developed to generate the initial population in the new environment.More specifically,decomposition-based interindividual correlation transfer learning learns the inter-individual correlation from the final population of the adjacent environment and transfers it to the new environment,aiming to maintain the diversity and distribution of the predicted population.While dimension-wise learning extracts the changing pattern of dynamic environments from the high-quality solutions of historical environments in the perspective of variable dimension to predict the high-quality genes,then the high-quality solutions are produced by splicing them,trying to improve the quality of the population and accelerate the convergence.Comprehensive experiments have been conducted by comparing the proposed method with four state-of-the-art algorithms on 14 benchmark problems.The results demonstrate that the proposed method can solve various dynamic multi-objective optimization problems stably and efficiently.