Research On Data-driven Multi-objective Evolutionary Optimization Method And Its Applications

In practical engineering applications,situations often arise where it is difficult to directly construct the objective function from first principles.In such cases,methods such as statistical learning or machine learning are frequently employed to fit sample data and construct surrogate models for objective prediction and optimization.These problems are referred to as data-driven optimization problems.To address such issues,it is first necessary to design and train surrogate models from sample data,and then obtain the optimal solution through optimization algorithms,thus obtaining the Pareto front that include conflicting criteria such as model complexity and model accuracy.This dissertation is aimed at tackling data-driven optimization problems by focusing on improving both the accuracy of surrogate models and the precision of optimization algorithms.Firstly,methods for enhancing the accuracy of surrogate models by utilizing coupling information between models are discussed and analyzed.Secondly,research is conducted on how to utilize local historical information in evolutionary algorithms to improve the precision of optimization algorithms.After decomposing these two aspects for research,a comprehensive discussion on data-driven optimization problems is carried out,with emphasis placed on research from the perspective of improving the utilization rate of raw data.Finally,the effectiveness of the proposed data-driven multi-objective evolutionary optimization method is validated through examples such as the optimization design of neodymium-iron-boron materials.Through this research,the aim is to provide effective methods and tools for solving complex optimization problems in practical engineering and contribute to the further development of the field of data-driven optimization.The content and contributions of the four works are summarized as follows:(1)Research on multi-conflict-expression symbolic regression method in datadriven modeling.Interpretability is required for data-driven models,but mainstream black-box machine learning methods make it challenging to interpret internal mechanisms,thereby increasing the challenge of interpretability.For multi-objective optimization,simultaneously optimizing multiple conflicting objectives is a common engineering problem,and building multiple surrogate models asynchronously leads to the loss of coupling information.Understanding the reasons for conflicts is also crucial for multi-objective engineering optimization solutions.Symbolic regression,as an interpretable modeling technique,is widely used because it can simultaneously explore model coefficients and structures.However,when facing multiple conflicting objective expressions,existing symbolic regression methods often require multiple asynchronous fittings,making it difficult to obtain coupling information,resulting in unsatisfactory results.A new multi-conflict expression symbolic regression algorithm is proposed to address this issue,and the concept of approximate maximum common subexpression is introduced to elucidate the reasons for conflicts.Additionally,an adaptive crossover matrix is added to balance the proportion of information exchange.Compared with single expression symbolic regression,the proposed method exhibits excellent performance.(2)Research on multi-objective data-driven evolutionary algorithm based on local ideal point set.Multi-objective evolutionary optimization algorithms face challenges in maintaining archives and selecting elite individuals when balancing convergence and diversity.A new multi-objective particle swarm optimization algorithm based on convergence contribution is proposed to address these issues.The algorithm uses a local ideal point strategy to evaluate convergence contribution and combines parallel cell distance to improve the convergence and diversity of the population.Unlike traditional methods,additional parameters or predefined reference points are not required by this algorithm.A set of local ideal points is built using information from the current archive during population evolution to calculate convergence contributions,thus selecting solutions with high convergence contributions as global best solutions.In the later stages of the algorithm,through the collaborative effect of parallel cell distance,the algorithm can maintain diversity of the archive while selecting global best solutions.Experimental results demonstrate the superiority of the proposed algorithm in solving test problems.(3)Research on offline data-driven evolutionary algorithm based on generative adversarial networks.In data-driven evolutionary optimization methods,ensuring the accuracy of optimization algorithms and the fitting accuracy of surrogate models is a key challenge.A data-driven multi-objective evolutionary algorithm assisted by generative adversarial networks is proposed in this research.The algorithm improves the accuracy of the optimization algorithm through critical fitness strategy and data augmentation strategy.The critical fitness strategy uses a discriminator to evaluate the prediction of surrogate models to ensure the reliability of the optimization process.The data augmentation strategy uses a generator to generate synthetic data to enhance the generalization ability of surrogate models.Experimental results verify the superiority of the proposed algorithm over other algorithms.(4)Application research of data-driven multi-objective evolutionary optimization.This study aims to verify the practical application effect of data-driven multi-objective evolutionary optimization methods in the engineering field,and takes the optimization design of Nd Fe B materials as a case study.In the optimization design of Nd Fe B materials,traditional methods have limitations due to the inability to describe the objective function with mathematical equations and the high experimental cost of materials.Through the data-driven multi-objective evolutionary optimization method,the heat treatment process parameters of magnetic material properties are successfully optimized,and good results are obtained in the experimental verification.The success of this case study verifies the potential of data-driven multi-objective evolutionary optimization methods in solving engineering problems,and provides effective tools and methodologies for engineering design.