摘要
基于Kriging代理模型的EGO(Efficient Global Optimization)算法的优化精度不高,并且当构建模型的样本数量较大时,算法优化变得耗时。对此提出了结合全局代理模型与局部代理模型的改进优化算法。使用Kriging作为全局代理模型,RBF作为局部代理模型,通过构建、优化多个局部代理模型来获取多个局部较优点,并在线更新Kriging模型,提高模型精度。针对优化耗时问题,提出了样本点遗忘法以及样本点渐进式增加法,使优化时间较EGO大大缩短。通过4个典型测试函数验证,并在收敛精度、稳定性两方面与EGO,PSO算法比较,结果显示两者都优于EGO与PSO,说明该算法具有强寻优性能、强鲁棒性。
The optimization accuracy of EGO (Efficient Global Optimization) algorithm based on Krigingsurrogate model is not high, and the number of sample points for constructing Kriging models is large, theoptimization is time consuming. This paper presents an improved algorithm combining global and localsurrogate models. Kriging is the global surrogate model, RBF (Radial Basis Function) is the local surrogatemodel, several local advantages points are getten by building, optimizing several local surrogate models, andthe Kriging model is updated online to improve the model accuracy. To solve the time-consuming problem ofmodel construction, this paper presents the sample point forgetting method and sample point incrementalmethod, greatly reducing the time of model building. Finally, four representative test functions are selected totest the performance of the algorithm and compare with EGO, PSO (Particle Swarm Optimization) in aspectsof convergence precision, stability, the result shows that the algorithm is better than EGO and PSO in twoaspects. The algorithm has strong optimization performance and robustness.
出处
《控制工程》
CSCD
北大核心
2017年第1期77-82,共6页
Control Engineering of China
基金
国家自然科学基金项目(61403140)
上海市自然科学基金(13ZR1411500)