Research On Dynamic Multi-objective Optimization Algorithm Based Transfer Learning

In production scheduling,artificial intelligence,combinatorial optimization,data mining,and many other areas of optimization,many complex dynamic optimization problems are often encountered,as they have multiple conflicting objectives that depend on changes in the environment,and the Pareto optimal solution changes with changes in the environment.As a result,optimization of dynamic problems can become difficult,with algorithms unable to adapt quickly to changes in the environment and unable to capture the full range of dynamic properties when faced with different types of dynamic problems.To address these problems,to make the designed method continuously adapt to the changing environment as much as possible,and to reuse valuable past knowledge to improve the speed and individual quality of solving different types of problems,this paper focuses on dynamic multi-objective optimization problems(DMOPs)using transfer learning strategies,aiming to explore the avoidance of negative transfer and solving different types of dynamic problems.This paper investigates three aspects of dynamic multi-objective optimization algorithms,namely,tracking historical information,predicting dynamic features from multiple perspectives,and dealing with different types of dynamic problems,respectively,and designs three different dynamic multi-objective optimization algorithms with the following main links:To address the problem of negative transfer in solving dynamic evolutionary problems using the strategy of transfer learning,this paper proposes a dynamic multi-objective evolutionary algorithm(T-DMOEA)based on streaming knowledge transfer to solve DMOPs.First,a multi-time prediction model is designed based on the movement trend of historical knee points,and a weighting method is used to effectively track the location of the knee points after environmental changes.Then,knowledge of sub-optimal solutions is reused in the set of non-knee points using manifold transfer learning techniques.The experimental results show that the proposed T-DMOEA algorithm combines knee points and high quality solutions to guide the generation of the initial population in the next environment during dynamic evolution.The algorithm is not only competitive in ensuring the diversity of the population,but also in avoiding the occurrence of negative transfer.In coping with the current dynamic multi-objective optimization methods that only solve the problem from a single space in the decision space or objective space,it has been to the extent that the algorithm is unable to obtain better estimates in the new environment.Therefore,a dynamic multi-objective evolutionary algorithm(M-DMOEA)based on multi-angle prediction is proposed in this paper.A kernel-based transfer learning model is designed to perform multi-angle prediction by considering both the decision space and the objective space at the same time.Experimental results show that the proposed design approach can significantly improve the performance of dynamic optimization.For dynamic optimization where there are different types of dynamic problems,it is important to respond effectively to changes in the environment.In this paper,an adaptive dynamic multi-objective optimization algorithm(TDA-DMOEA)based on type detection is proposed,while the dynamic detection operator designed in this strategy is used to identify the types of dynamic problems.Different types of dynamic changes are also given different response strategies.A comprehensive study of the commonly used DMOPs benchmark test function set is carried out,and experimental results show the effectiveness of the method.