Abstract: For the contradiction between time and performance in the modeling of combined-kernel support vector machines，it proposes a new scheme in finding the ideal compromise of performance and time.Kernels’combination is studied beneficial for learning and generalization ability，as well as the method of optimizing kernels’parameters.A master-slave kernel phase-optimized programs is proposed，namely a time optimizes parameters of one kernel，and gradually adds other sub-kernel.Time cost is the simple sum of time every single-kernel modeling consumed.Compared with the evolutionary algorithm，time consumed is less，and compared with the partition algorithm，the performance is better.The proposed scheme in terms of time and performance achieves good results.

Key words: Fisher, combined kernel, multi-kernel, kernel parameters, support vector machines

摘要： 针对组合核支持向量机建模中存在的耗时和性能的矛盾问题，提出新的方案，用于在时间和性能上寻找理想折衷。研究了兼顾学习和推广能力的核组合，以及优化核参数方法。提出了一种主从核逐步优化的方案，即每次只优化一个核的核参数，逐步加入其他子核求解参数，时间上大致是求解单核参数耗时的简单叠加，相对于进化算法求解模型耗时更少，相对于分治算法求解模型性能更优。提出的方案在时间和性能上取得了较好的效果。

关键词: Fisher准则, 组合核, 多核, 核参数, 支持向量机