Abstract: A fusion kernel optimization algorithm based on the fusion kernel is presented, which improves the method by Xiong et. al. The proposed method uses a data-dependent kernel to maximize the kernel Fisher criterion, so that the different kernels can be fused. Then the method is employed to optimize the kernel of KPCA and the evaluation of the classification performance based on the toy data and measured radar high range resolution profile (HRRP) data show the greater efficiency of our method.

Key words: kernel machines, class separability, empirical feature space, kernel optimization, high-resolution range profile(HRRP)