第5卷第2期 智能系统学报 Vol.5 No.2 2010年4月 CAAI Transactions on Intelligent Systems Apr.2010 doi:10.3969/j.issn.16734785.2010.02.008 SVM性能的免疫鱼群多目标优化研究 刘胜,李高云12,江娜 (1.哈尔滨工程大学自动化学院,黑龙江哈尔滨150001;2.中国船舶重工集团公司第七0七研究所九江分部,江西 九江332007) 摘要:SVM算法的训练精度和训练速度是衡量其性能的2个重要指标.以这2个指标为目标变量建立SVM性能多 目标优化问题的数学模型,采用直接对多个目标同时进行优化的方法求得问题的Pareto近似解集.在求解Pa©to近 似解集时,将免疫原理中的浓度机制引入基本鱼群算法中,形成一种改进的免疫鱼群算法.以非线性动态系统仿真 数据为样本数据,并采用改进的免疫鱼群算法求解SVM性能多目标优化问题的Pareto近似解集.仿真结果表明,在 解决多目标优化问题时,免疫鱼群算法相对于基本鱼群算法和遗传算法具有更好的优越性。 关键词:支持向量机;多目标优化;Pareto近似解集;免疫鱼群算法 中图分类号:TP183文献标识码:A文章编号:16734785(2010)02-0144-06 Multi-objective optimization of an immune fish swarm algorithm to improve support vector machine performance LIU Sheng',LI Gao-yun'2,JIANG Na' (1.College of Automation,Harbin Engineering University,Harbin 150001,China;2.Jiujiang Branch of 707 Research Institute,China Shipbuilding Industry Corporation,Jiujiang 332007,China) Abstract:Accuracy and speed when training a support vector machine (SVM)algorithm provides critical measure- ments of the algorithms performance.To optimize performance,a mathematical model of multi-objective optimiza- tion with improvements in these two parameters as goals was established.A Pareto approximate solution set was ob- tained by optimizing multiple targets simultaneously.In the process of finding the Pareto approximate solution set,a concentration mechanism from an immune algorithm was introduced into the basic artificial fish swarm algorithm. This produced significant improvements and resulted in the proposed immune fish swarm algorithm.Taking the non- linear dynamic system simulation data as sample data,a Pareto approximate solution set of multi-objective optimiza- tion of SVM performance was obtained using the improved algorithm.Simulation results showed that,for solving multi-objective optimization,the immune fish swarm algorithm was superior to both a basic artificial fish swarm al- gorithm and to genetic algorithms. Keywords:support vector machines;multi-objective optimization;Pareto approximate solution set;immune fish swarm algorithm 支持向量机(support vector machine,SVM)Iu2] 目前,MO0问题的求解方法可以大致分为2类,一 算法的训练精度是最受关注的性能指标之一,同时 类是将MO0问题转化为单目标优化问题进行求 训练速度也是一个关键因素,尤其在学习速度要求 解,这种方法最终只能求出一个综合了各个目标分 较高的场合.本文综合考虑算法的训练精度和速度, 量信息的最优解,但各个目标的重要程度是在构造 从多目标优化(muli-objective optimization,MO0)的 相应的单目标问题时确定的,不具有普遍性.另一类 角度出发,对SVM回归算法的2个指标进行研究和 方法是直接对多个目标同时进行优化,即多目标问 探讨.Schaffer首次采用进化算法研究了MOO问题. 题的直接求解方法.MOO直接求解法多数为智能类 收稿日期:20090324. 算法,都是同时求出多目标规划问题的Pareto最优 基金项目:黑龙江省自然科学基金资助项目(A2004-19). 解集.Shelokar等人[3]提出了一种基于Pareto支配 通信作者:刘胜.E-mail:iu.sch@163.com