Research On Particle Swarm Optimization Based On Gaussian Process Surrogate Model

Swarm intelligent computation is an important optimization method.It has a good robustness and parallelism,and it can solve robust,discontinuous and non-differential black-box optimization problems,which makes it a hot topic in recent research.As an iterative algorithm,swarm intelligence algorithm often needs a lot of fitness evaluations.In solving practical problems,the calculation of fitness often consumes a great deal of time cost,which seriously affects the efficiency of the swarm intelligence algorithm.The surrogate model can be built to replace the complex function model according to the date received.Therefore,based on the data-driven technology,particle swarm optimization(PSO)is used as the optimal search algorithm in this dissertation;various efficient surrogate models are constructed to evaluate the fitness of the particles.The main contents are as follows:1.A cooperative particle swarm optimization algorithm based on master-slave Gaussian process model is proposed.When the surrogate model is used to evaluate the particles,the master Gaussian process model and the slave Gaussian process model are used to evaluate the search population respectively.In this paper,the master Gaussian process model is used to evaluate the particles from the global point of view,and the quadratic evaluation is carried out at each local optimum by the slave model.The cooperation of the two models greatly improves the ability ofthe algorithm.2.An adaptive surrogate model particle swarm optimization based on EI(Expected Improvement)Sampling is proposed.Considering that the frequent updating of the surrogate model in the process of model management takes a certain amount of computing time and to avoid invalid and time-consuming model updating,an adaptive updating method of surrogate model is proposed,and the accuracy of the model is improved by combining an EI sampling method.3.A social learning particle swarm optimization algorithm based on ensemble model is proposed.In this algorithm,several basic surrogate models are weighted to form the ensemble model to reduce the prediction error of the model.In this paper,the isomorphism ensemble model and heterogeneous ensemble model are proposed and the optimization performance of the two models is analyzed.Finally,the social learning particle swarm optimization is combined with the ensemble model to improve the searching ability of the algorithm and the ability to solve practical problems.The simulation tests and performance analysis of the above methods are carried out in this dissertation with several benchmark numerical optimization problems and the spring structure design problem.The results show that the proposed methods can improve the optimization performance of the algorithm to some extent,and it can effectively makea balance between the evaluation cost of complex problems and the optimization efficiency.So it can be known that using the data-driven technology to build the surrogate model is an effective way to improve the performance of swarm intelligent computing,and it can effectively improve the ability to solve practical problems.