A relevance vector regression based metamodeling approach for complex system analysis

The metamodeling approach has been an important method to reduce the computational expense of complex system simulation. Metamodeling is the process of building a ldquomodel of a modelrdquo to provide a fast surrogate model for computational expensive simulation code. Main metamodeling techniques include polynomial regression, kriging, radial basis function and support vector regression. In this paper we investigate relevance vector regression (RVR) as an alternative metamodeling approach for complex system simulation. To further understand this new method, we compare its performance with other four metamodeling method using test functions. RVR achieves more accuracy than four other metamodeling approaches and have good robustness and acceptable computational efficiency. The results suggest the RVR approach has powerful potential for metamodeling applications.