Research On Offspring Generation Method Of Multi-objective Evolutionary Algorithms Based On Clustering

Multi-objective optimization problems with multiple constraints and variables are widely found in scientific research and engineering practice.Multi-objective evolutionary algorithms based on evolutionary theory do not require a priori information about the problem,which perform well in solving multi-objective optimization problems by good parallelism and robustness.The generation of offspring is an important part of the multiobjective evolutionary algorithm and significantly affects the performance of the algorithm.Therefore,it is important to study the generation methods of offspring of multi-objective evolutionary algorithms.In the multi-objective evolutionary algorithm,offsprings are generated through two parts: parent selection and individual recombination.Improving the way of random parent selection in multi-objective evolutionary algorithms helps to avoid misleading the convergence direction of the algorithm.Improving the recombination operator of multiobjective evolutionary algorithm helps to improve the quality of offspring individuals and enhance the performance of the multi-objective evolutionary algorithm.We use K-means clustering algorithm to mine the information of the population.Mating restriction on parent selection based on the clustering results is designed to balance the global search and local search.Different recombination operators are designed to guide the algorithm to balance diversity and convergence.The improved algorithm is applied to solve the PV maximum power point tracking problem.The main research contents and results are as follows.A clustering-based dynamic mating restriction probability strategy(CDMRS)is proposed.The CDMRS uses the K-means algorithm to divide the population,which can effectively control the origin of the parents.Based on the clustering results,the mating restriction probability is dynamically updated in different stages of the algorithm according to the search demand.On 42 standard test problems,the improved algorithm obtained better IGD and HV metrics on 30 and 35 problems respectively by embedding CDMRS into NSGA-II.The improved algorithm obtained better IGD and HV metrics on 23 problems by embedding CDMRS into SPEA2.The effectiveness and adaptability of the strategy were verified.A two-stage reorganization operator multi-objective evolutionary algorithm based on clustering(TROC-MOEA)is proposed.The model-based recombination operator is designed to construct a Gaussian model for each individual in class based on the clustering results.The local mining ability of the algorithm is improved by sampling to generate offspring.A two-stage strategy is designed to use genetic operators or model operators at different stages of algorithm evolution with preferences for exploration and exploitation,to balance global search and local search.TROC-MOEA and five classical algorithms are tested on 42 standard test problems,and the optimal IGD and HV metrics were obtained on21 and 19 problems respectively,verifying the effectiveness and rationality.TROC-MOEA is applied to solve the PV maximum power point tracking problem.The PV maximum power point tracking problem is modeled as a multi-objective optimization problem of minimizing the current error and maximizing the output power.TROC-MOEA and five classical algorithms are experimented under four different simulation environments designed,TROC-MOEA obtains larger output power and output voltage with smaller current error.