基于自适应均值粒子群算法的SVM参数优化方法

针对传统粒子群算法(Particle Swarm Optimization,PSO)对支持向量机(Support Vector Machine,SVM)参数寻优时的低效问题,运用了自适应均值粒子群算法(Adaptive Mean Particle Swarm Optimization,MAPSO)对SVM参数进行优化(MAPSO-SVM算法).采用自适应策略,引入了余弦函数、非线性动态调整惯性因子,每次进化都根据种群中粒子的适应度值大小将粒子分为3个等级,对每个等级的粒子赋予相应的惯性因子,将PSO算法速度更新方程中的个体历史最优位置和全局最优位置用它们的线性组合代替.分别用SVM、PSO-SVM和MAPSO-SVM算法对UCI中不同数据集进行实验测试,结果表明MAPSO-SVM算法比SVM和PSO-SVM算法的分类效果更好,分类准确率比SVM和PSO-SVM算法分别平均提高了14.7290％和1.8347％,同时与PSO-SVM算法相比,算法的收敛精度和效率更高.

Parameter Optimization of SVM Based on Adaptive Mean Particle Swarm Optimization

In order to solve the inefficiency problem of parameter optimization of support vector machine (SVM) with the particle swarm optimization (PSO) algorithm,the adaptive mean particle swarm optimization (MAPSO) algorithm is used to optimize the SVM parameters (MAPSO-SVM algorithm).The adaptive strategy was used,and the cosine function and nonlinear dynamic adjustment inertia factor were introduced.The particles were divided into three grades according to the fitness value of the particles in the population each evolution,and the corresponding inertia factor was given to each grade particle.The individual optimal history position and the global optimal position in the PSO algorithm speed update equation were replaced by their linear combinations.The experimental results of SVM,PSO-SVM and MAPSO-SVM for different datasets in UCI show that the MAPS0-SVM algorithm is better than SVM and PSO-SVM algorithm in classification,and the classification accuracy is 14.7290％ higher than that of SVM and 1.8347％ higher than that of PSO-SVM algorithm.Compared with PSO-SVM algorithm,the convergence accuracy and efficiency of the algorithm are higher.