New heuristic for determination Gaussian kernels parameter

For SVM using Gaussian kernel, it was known that parameter σ changing from infinity to zero corresponds SVM model performing severe under-fitting to severe over-fitting. In this paper, we theoretically investigated the correspondence between Gaussian kernel´s parameter and support vectors´ distribution, interesting result was found: with σ´s changing from big to small, the samples nearest to other class first become support vectors, then nearest and farthest ones, finally all samples become support vectors. Then we proposed a new heuristic algorithm based on Zhangling et al.´s boundary points theory for selecting Gaussian kernel´s parameter. Experimental results showed this heuristic performed well on selecting a suitable a for Gaussian kernel.