基于相关向量回归的仿真元建模方法  

Simulation Metamodeling Approach Based on Relevance Vector Regression

作　　者：吴冰[1] 程思微[1] 张文琼 梁加红[1] 

机构地区：[1]国防科技大学机电工程与自动化学院,长沙410073 [2]第二炮兵网管中心,北京100085

出　　处：《计算机工程》2010年第3期24-27,共4页Computer Engineering

基　　金：国家自然科学基金资助项目(60704038)

摘　　要：针对支持向量回归元模型存在的不足,提出将相关向量回归应用于仿真元建模,使用多个不同维度和非线性程度的基准测试函数,在元模型精确性、采样技术、样本规模、模型维度和非线性程度等多方面与多项式回归、Kriging、径向基函数、支持向量回归4种方法进行对比研究,结果证明该方法具有较高的精确性和鲁棒性。Aiming at the shortcoming of support vector regression metamodel, Relevance Vector Regression（RVR） is investigated as an alternative metamodeling approach. Using several benchmark test functions with varying model dimensions and degrees of nonlinearity, RVR is compared with four metarnodeling approaches, including polynomial regression, Kriging, radial basis function and support vector regression. Several performance criterions are considered, including metamodel accuracy, effect of sampling techniques, effect of sample size, effect of model dimensions and degrees of nonlinearity. Results prove that RVR approach can achieve higher accuracy and more robustness.

关 键 词：元建模 相关向量回归 支持向量回归 

分 类 号：N945.12[自然科学总论—系统科学]