Abstract: In the process of parameter optimization for support vector machines ( SVMs) with Gaussian kernel,in- ter-cluster distance in feature spaces ( ICDF) is an effective measure.However,ICDF may result in heavy compu- tational load and large time consumption.In order to solve this problem,firstly,the theorem that ICDF is a positive strictly-unimodal function about Gaussian kernel parameter is proved.Then,according to this theorem,a modified golden section algorithm ( MGSA) is proposed to search a shrunk value fast for kernel parameter in the candidate set.Thus,a fast parameter optimization approach on the basis of both MGSA and differential evolutionary algorithm is presented.Finally,some experiments are carried out to verify the effectiveness and rapidity of the proposed ap- proach.

Key words: support vector machine, inter-cluster distance, parameter optimization, kernel parameter, modified golden section algorithm