基于矢量基学习的自适应迭代最小二乘支持向量机回归算法

Adaptive Iterative LS-SVM Regression Algorithm Based on Vector Base Learning

为增强最小二乘支持向量机（LS-SVM）回归建模的稀疏性、鲁棒性和实时性,在加权LS-SVM的基础上,提出了基于矢量基学习的自适应迭代回归算法。在训练过程中,该算法通过矢量基学习和自适应迭代相结合的方法得到1个小的支持向量集,同时采用加权方法确定权值以减小训练样本中非高斯噪声的影响。回归学习和动态系统辩识的仿真结果表明：在回归建模精度相似的情况下,该算法确定的支持向量为全部学习样本的4.9%~8.9%,训练时间为标准LS-SVM的0.011%~0.383%;由于能够鲁棒跟踪时变非线性系统的动态特性,适合在线实时训练;可进一步用于非线性系统的建模和实时控制研究。

To enhance the sparseness,robustness and real-time performance for regression modeling of least square support vector machine（LS-SVM）,an adaptive iterative regression algorithm based on the vector base learning and weighted LS-SVM is proposed.In the algorithm ＇ s training process,the vector base learning and adaptive iterative procedures are combined,and a small support vector set can be obtained.The weights are determined by a weighted method in order to reduce the effect of the non-Gaussian noise in training samples.Simulation results of regression learning and dynamic system identification show that under the condition of similar regression modeling accuracy,the support vectors that the proposed algorithm determines are 4.9%~8.9% of the whole training samples,and the running time is 0.011%~0.383% of the standard LS-SVM.The algorithm is suitable for on-line real-time training because it can robustly track the dynamic characteristics of nonlinear time-varying systems and can be applied in nonlinear systems modeling and real-time control.