Research On Dynamic Multi-objective Optimization Based On Evolutionary Algorithms

Multi-objective optimization has attracted more and more attention in recent years.Applying evolutionary computation to multi-objective optimization is a very effective method to solve multi-objective optimization problems.In addition to static multi-objective optimization problems,dynamic multi-objective optimization problems(DMOPs)have also attracted many people's attention.The objective functions in dynamic multi-objective optimization problems conflict with each other and change with time.To solve this problem,many dynamic multi-objective evolutionary algorithms(DMOEAs)have been proposed.These DMOEAs combine traditional evolutionary algorithms with some dynamic mechanisms,including diversity introduction methods and prediction methods.However,most DMOEAs can only adapt to a fixed type of dynamic environment.In order to improve the performance of existing dynamic multi-objective optimization algorithms,two dynamic multi-objective evolutionary algorithms are proposed from different perspectives.(1)A dynamic multi-objective evolutionary algorithm based on decision variable classification(DMOEA-DVC)is proposed.The algorithm divides static decision variables into two categories and dynamic decision variables into three categories.In static optimization,different crossover operators are used for different static decision variables,so that the algorithm can accelerate convergence while maintaining diversity.In dynamic optimization,DMOEA-DVC introduces a hybrid change response strategy.Once environmental change occurs,DMOEA-DVC will reinitialize the three dynamic decision variables by maintenance,prediction and diversity introduction respectively.DMOEA-DVC is compared with eight most advanced algorithms on 33 benchmark problems with different characteristics.Theexperimental results show that the overall performance of DMOEA-DVC is better than that of the comparative algorithms.(2)A dynamic multi-objective evolutionary algorithm based on population adaptability(DMOEA-PA)is proposed,which can adapt to various dynamic multi-objective problems adaptively.First,DMOEA-PA applies inverse generation distance(IGD)in decision space and evaluates the quality of trial populations generated by different strategies by IGD.This evaluation method does not need to consume the number of objective function evaluations or the real Pareto optimal front.Then,through the quality of the experimental population,we can find out the corresponding generation strategy of the experimental population which is more suitable for the current environment.Thus,the probability of obtaining high quality population is greatly improved.Whether it is the generation strategy in static optimization or the population re-initialization strategy in response,the above methods can be used to find the best generation strategy.DMOEA-PA is compared with 8 state-of-the-art algorithms on 33 benchmark problems with different characteristics.The experimental results show that the overall performance of DMOEA-PA is better than that of the comparative algorithm.