Research On The Application Of Bayesian Evolutionary Optimization To High Dimensional Expensive Many-objective Problems

The "Made in China 2025" released in 2015 pointed out that it emphasizes the acceleration of generation information technology and production technology to promote intelligent optimization control of production processes.The multi-objective optimization problem is a typical research problem of intelligent optimization control.Bayesian evolutionary optimization algorithm has been widely used to solve multi-objective optimization problems which are computationally expensive.However,most existing Bayesian evolutionary optimization algorithms are only suitable for low dimensional problems.On the one hand,the uncertainty measures for different individual predicted values lost significant differences.On the other hand,as the data becomes sparse,it is difficult for the population to search the solution space effectively.And individuals with more potential cannot be selected for expensive calculation.The study improved the Bayesian evolutionary optimization algorithm and designed a variety of strategies to improve the algorithm performance for the high-dimensional expensive multi-objective optimization problem.The main contributions and achievements are as follows:1.In order to solve the optimization bottleneck caused by dimensionality disaster in high-dimensional expensive multi-objective optimization problem,Bayesian co-evolutionary optimization algorithm is proposed based on information entropy.In the algorithm,the information entropy model is introduced into the co-evolutionary algorithm.The whole population is assigned an uncertainty measure based on the entropy value.In order to ensure the convergence rate of populations,excellent individuals are exchanged by the non-dominated sorting and information entropy models.And individuals are searched in less developed areas to increase population diversity.The experimental results show that compared with the existing algorithms.The proposed algorithm has certain competitiveness in the convergence,diversity and distribution of the population.2.In the high-dimensional space,the reliability of the prediction information is reduced.The prediction performance of the addition criterion is poor.And the effective prediction samples can not be selected to guide the population to evolve in the correct optimization direction.The thesis proposes a two-stage Bayesian evolutionary optimization algorithm based on generation adversarial networks to achieve expensive multi-objective optimization.The generation adversarial network and co-evolution are combined.And the prediction samples that are closer to the real samples are selected in the environmental selection as the parent-guided population evolution.It reduces the deviation of the optimal direction due to the prediction error.At the same time,the algorithm combines Lp-norm form with the maximum and minimum distance criterion.The strategy can effectively reduce the complexity of calculation,solve the problem of sparse samples in the case of small samples and select more promising individuals for the actual calculation.3.On the basis of the model and the optimization algorithm,the application verification study is carried out with the background of neural architecture search.Firstly,the experimental parameters are analyzed based on the actual data generated by the improved algorithm on the test dataset.And it is compared with other high-performance optimization algorithms to prove the performance of the improved algorithm when solving the problem in the study.