Research On Online Support Vector Machine And Imbalanced Learning Based Dynamic Multi-Objective Optimization Algorithms

The Dynamic Multi-objective Optimization Problems exist in lots of aspects of our life,but because it has multiple conflicting goals that depend on time at the same time,and its optimal functions will changes with time,it makes the Dynamic Multi-objective Optimization Problems are very difficult to solve.Solving the Dynamic Multi-objective Optimization Problems have important application and theoretical value.In the study of dynamic multiobjective optimization problems,evolutionary algorithms have shown great advantages and attracted the attention of researchers at home and abroad.However,most of the existing methods ignore the importance of the initial population to the Dynamic Multi-objective Evolutionary Algorithms.So the main idea of this paper is to use efficient and stable means to use the historical optimal solution information to find a better initial population,so as to solve the dynamic multi-objective optimization problem more efficiently.For the Dynamic Multi-objective Optimization Problems,this paper first proposes a SVM based Dynamic Multi-objective Optimization Algorithm SVM-DMOEA.We use the solutions in the POS obtained at the historical moment to train the SVM classifier,and then use this classifier to get the initial population at the future moment.The advantage of this algorithm is to combine mature machine learning technology with classic Multi-Objective Optimization Evolutionary Algorithms.This method can retain the global search ability of Evolutionary Algorithms and the more mature classification ability advantage of SVM.Then,a Dynamic Multi-objective Optimization Algorithm based on Incremental SVM called ISVM-DMOEA was proposed.This algorithm uses the POS information of historical moments to perform online training of SVM classifiers in an online manner,to generate a good initial population in the future.The advantages of this algorithm:the idea of online learning is introduced into the process of solving DMOP,and the SVM classifier is trained online,which saves the calculation of invalid repeated historical information and saves computing resources more effectively.Finally,on the basis of ISVM-DMOEA,we propose a Dynamic Multi-Objective Optimization Algorithm SMOTEISVM-DMOEA based on an imbalanced online support vector machine.This algorithm has the following advantages:Firstly,the online learning method can efficiently reuse the information from the past moments to obtain a prediction model,which can improve the search accuracy of DMOP at other times.Secondly,the Imbalanced data learning method further improves the performance of the ISVM classifier,which can more effectively save the loss of computing resources and better respond to real-world applications.Finally,we prove the effectiveness of the algorithm through experimental results.The significance of this research lies in the in-depth study of the dynamic multi-objective optimization problem,which combines machine learning,online learning and Imbalanced data methods step by step,and proposes three effective algorithms.This not only improves the solution performance of Dynamic Multi-objective Optimization problems,but also is helpful for related research on Dynamic Multi-objective Optimization Problems,and has a profound impact on the practical application of Dynamic Multi-objective Optimization Problems.