摘要
针对支持向量数据描述(SVDD)训练过程中的参数优化问题,提出了一种有限穷举—局部遗传算法.首先,在分别分析参数C和σ对SVDD分类性能不同影响的基础上,得到参数σ是影响分类性能主因的结论.然后针对σ的优化问题,通过穷举有限个整数解并比较其分类性能来确定近似最优解,在近似最优解的领域内用遗传算法进行局部搜索,最终得到精确的优化参数.仿真实验及电路故障检测应用结果表明:算法有效避免了参数搜索的盲目性,能以更短的时耗逼近最优解.
Aiming at the problem of SVDD parameter optimization in training process, an finite exhaustion-local genetic algorithm is proposed. Based on the analysis of C and σ parameters＇ different influence on SVDD classfication performance respectively, a conclusion that σ is the main influencing factor is made. Focus on optimization issue, the near--optimum solution of σ is determined by enumerating finite integer solutions along with comparing these solutions＇ performance, and then search is launched locally in nearby domain of near--optimum solution. The precise parameter is get at last. Simulation experiment and circuit fault detection application results show that the above algorithm avoids blind parameter searching and can approach optimum solution in shorter time wastes.
出处
《微电子学与计算机》
CSCD
北大核心
2013年第2期13-17,共5页
Microelectronics & Computer
基金
河北省重点基础研究项目(10963529D)
