Research On Hybrid Multi-strategy For Dynamic Multi-objective Evolutionary Optimization

Dynamic multi-objective optimization problems have multiple con-flicting objectives,constraints,and the related parameters,etc.,which may change over time.When solving dynamic multi-objective optimization problems,it is often required that the algorithm can find its Pareto optimal set before the problem changes.The solution methods often used by researchers are diversity-based strategies,memory-based strategies,and prediction-based strategies.However,these methods have some weaknesses: the diversity-based strategy can ensure the global search ability of the algorithm but cannot accelerate the population convergence;the memory-based strat-egy is suitable for periodic changes,and for non-periodic changes,may mislead the search direction of the population;the prediction-based strategy learns the historical population search experience and predicts the location of the population in the new environment,and its effectiveness depends on the validity of the model.This study considers the performance differences of diversity strategy,memory strategy and pre-diction strategy from the characteristics of dynamic multi-objective optimization,and uses the advantages of each strategy to design algorithms based on multi-strategy fu-sion.(1)A Transfer Learning-based Knee Point Prediction Strategy for Dynamic Multi-objective Optimization is proposed,called TKPS.TKPS uses a prediction model based on transfer learning to learn the historical population search experience and predict the location of new populations by using transfer learning methods.TKPS uses a memory strategy to keep the historical optimal solutions in a memory pool.When the environ-ment changes,in order to improve the effect of the transfer learning algorithm,TKPS selects the appropriate individuals from the memory pool for the transfer learning al-gorithm.In order to give full play to the advantages of memory strategy,tkps selects individuals with high fitness from the memory pool to join the new species group to speed up the convergence of the population.In addition,to increase the population di-versity and improve the global search ability of the population,TKPS generates some individuals randomly in the decision space to join the new population.Finally,TKPS is compared with other three optimization algorithms.The experimental results show that the multiple strategies integrated by TKPS can effectively solve most of the tested functions.(2)A Dynamic Characteristic-Based Multi-Model Prediction Method for Dy-namic Multi-objective Optimization is proposed,called DCMM.DCMM designs var-ious prediction strategies based on the characteristics of environmental changes in dynamic multi-objective optimization problems.DCMM uses a memory strategy to preserve the centroids and inflection sets of historical populations.When the envi-ronmental change is similar to the historical change,the centroid and inflection point sets of the historical population are used to guide the population evolution;otherwise,DCMM establishes whether the Pareto optimal solution set has changed.If the Pareto optimal solution set changes,individuals of the population in the new environment are generated through a set of knee points based on the two previous consecutive popula-tions.If not,new populations are generated through the perturbation-based approach.In addition,DCMM adaptively adjusts population diversity based on the intensity of environmental changes to ensure the global search capability of the algorithm.Finally,DCMM is compared with other advanced optimization algorithms.The experimen-tal results show that the strategy integration method designed by DCMM based on dynamic change characteristics can improve the convergence and distribution of the algorithm.