一种改进粒子群算法的混合核ε-SVM参数优化及应用

Parameters optimization and implementation of mixed kernels ε-SVM based on improved PSO algorithm

将径向基核函数和多项式核函数进行线性组合构建了混合核ε-SVM,克服了单核SVM存在的泛化性能弱、学习能力差等弱点;为了同时解决普通粒子群算法存在的后期震荡严重、趋同性强和极易陷入局部极小值等问题,提出了一种改进的PSO算法,并给出了其数学模型和算法流程。该算法将随机粒子个体极值的追随因子增加至动量项和基本粒子群算法的速度项,再将增加追随因子后的动量项回植于更新后的速度项,这样就使得粒子在减缓后期震荡的同时修正了趋同性。通过函数仿真实验和实例验证了所提出的基于改进PSO的混合核ε-SVM算法较其他预测算法具有寻优精度高、收敛速度快、鲁棒性能好和复杂度低等优势。

This paper constructed a mixed-kernel ε-SVM by combining RBF kernel and polynomial kernel linearly in order to use SVM to predict more effectively and overcome some typical shortcomings（like weak generalized performance and learning capability of normal SVM）.As there were many problems existing in standard PSO in the process of the latter stage,such as serious concussion,inclination-tendency and the high possibility of being involved in local maximum value,this paper proposed an improved PSO algorithm to solve the above problems at the same time as well as its mathematic model,and gave algorithmic procedure.It imported the following-factor of random-particle＇s maximum value to the expressions of momentum and the velocity of normal PSO algorithm,then back embedded the new momentum expression to the newly updated velocity formular,by which particle could weaken the concussion and inclination tendency simultaneously.A function simulation and a real-data based experiment prove that the mixed kernel ε-SVM based on improved PSO presented in this paper has good advantages over many other predicting algorithms in forecasting precision,convergence velocity,robustness and simplicity,therefore,it＇s a valuable regression algorithm to be spread.