Modified Particle Swarm Optimization for Multi-Scale Kernel Function in SVM

Multi-scale kernel function learning is a special case of multi-kernel learning, namely combines several multi-scale kernels. This approach is more flexible. It provides more comprehensive choice of scale than the mixed kernel learning. In this paper, the model´s parameters of multi-scale Gaussian kernel were used as elementary particles. The parameters of multi-scale Gaussian kernel were global optimized by modified particle swarm optimization algorithm (PSO). Finally the optimal prediction results can be found. Experimental results show that: modified PSO can much more preferably optimize the parameters of multi-scale kernel model. It influences and improves the performance of multi-scale kernel and enhances the final accuracy of classification.