Global Simulation And Optimization Algorithm Based On Support Vector Regression Model

Traditional global optimization methods and other simulation and optimizationmethods based on alternative models have shortcomings that they need many evaluationtimes and can not cope with high-dimensional optimization problems and so on. Themethod based on the "black box" meta-model (response surface) became popular inrecent years. It mainly includs global optimization methods which based on SVR, RSM,Kriging, RBF and other meta-models.This method is a kind of global optimizationmethod based on experimental design and mathematical statistics. It obtains a sufficientlyaccurate function between the design variables and design objectives by lessexperiments.It reduces the calculation cost of optimization problem by alternative modelof the response surface effectively.Support vector regression (SVR, Support Vector Regression) based on SVMestablishes the separating hyperplane to construct an alternative response surface modelof source model by optaining maximum interval of training samples. Existing globaloptimization methods based on SVR can not guarantee that it can select a representativesample to cover the entire design range when the simulation sample size is small;reconstruction SVR model need a long time; searching the optimal point is slow; and cannot effectively deal with constrained global optimization problems.This paper proposed a new improved global optimization algorithm DISVR based onincremental SVR model: an incremental SVR method to improve the reconstructionefficiency of processing the response surface, a new incremental LHD sampling（LatinHyper-cube Sampling）to ensure an uniform distribution of samples; DIRECT searchalgorithm to enhance the stability and efficiency of the global search.This paper discoursed the important parts of DISVR in the relevant sections, The testresults of Standard and Project showed that the proposed DISVR algorithm canefficiently and stabilizilly get the optimal results, and had a good prospect.