LS-SVM参数估计与稀疏化方法研究及应用

Method for Parameters Estimation and Sparseness of LS-SVM and Application

针对最小二乘支持向量机(LS-SVM)参数估计与稀疏化问题,提出了采用多智能体遗传算法(MAGA)估计LS-SVM参数并在参数估计的过程中实现稀疏化的方法。首先,根据被估参数的特点,设计了MAGA的种群初始化方法和各遗传算子操作方式。其次,基于剪枝法的思想,设计了包含三种控制条件的稀疏化策略,能够在不明显降低回归精度的前提下实现LS-SVM的稀疏化。最后,通过实例计算验证本文方法的有效性,计算结果表明,MAGA较其它方法(自适应遗传算法、粒子群算法)能够获得更优的参数,从而使LS-SVM具有更优的回归性能,并且稀疏化策略稳定有效。

A method for estimating parameters and realizing sparseness of least squares support vector machines（LS-SVM） using multi-agent genetic algorithm（MAGA） was proposed. Firstly, population initialization method and genetic operators＇ operation mode of MAGA were designed according to the characteristics of parameters to be estimated. Secondly, the sparseness strategy including three control conditions was designed based on the idea of pruning algorithm, which could realize sparseness of LS-SVM without lowering regression accuracy visibly. Lastly, validity of method was verified through calculational examples. The results indicate that the LS-SVM has better regression performance than others because MAGA can get more optimal parameters, and the sparseness strategy is stable and effective.