首先需要下载LSSVM工具箱,设置添加文件夹路径到matlab中;
主要任务:PSO优化LSSVM的两个参数gam,sig2.
main.m代码:
[train_x,train_y] = loadtraindata();%一次性导入训练数据
P_number=20;%粒子群个数
C1=2;%初始化学习因子
C2=2;
W_max=0.95;%初始权重
W_min=0.4;%终止权重
K=100;%迭代次数
gammax=300;%定义优化参数的取值范围
gammin=0.01;
sig2max=200;
sig2min=0.01;
V_max=25;%定义速度取值
V_min=-V_max;
P_xy=zeros(P_number,3);%矩阵方式定义微粒的位置和速度
P_v=zeros(P_number,2);
Pbest=zeros(P_number,3);
gbest=zeros(1,3);
%矩阵方式定义每个微粒的最好位置以及每个最优适应度
for i=1:P_number
Pbest(i,1)=(gammax-gammin)*rand+gammin;
P_xy(i,1)=Pbest(i,1);%初始化微粒信息
Pbest(i,2)=(sig2max-sig2min)*rand+sig2min;
P_xy(i,2)=Pbest(i,2);
Pbest(i,3)=adapt(Pbest(i,1),Pbest(i,2),train_x,train_y);
P_xy(i,3)=Pbest(i,3);
P_v(i,1)=V_max*rands(1,1);%初始化微粒的随机速度信息
P_v(i,2)=V_max*rands(1,1);
end
%矩阵方式定义所有微粒的最好位置以及全局最优适应度
gbest(1,1)=(gammax-gammin)*rand+gammin;
gbest(1,2)=(sig2max-sig2min)*rand+sig2min;
gbest(1,3)=adapt(gbest(1,1),gbest(1,2),train_x,train_y);
for j=1:K
for i=1:P_number
i=1;
gam=P_xy(i,1);
sig2=P_xy(i,2);
ad=adapt(gam,sig2,train_x,train_y);%评价每个微粒的适应度
P_xy(i,3)=ad;
if adV_max)
P_v(i,1)=V_max;
end
if(P_v(i,1)V_max)
P_v(i,2)=V_max;
end
if(P_v(i,2)gammax)
P_xy(i,1)=gammax;
end
if(P_xy(i,1)sig2max)
P_xy(i,2)=sig2max;
end
if(P_xy(i,2)
loadtraindata.m代码(导入训练数据集):
function [train_x,train_y]=loadtraindata()
train_x=importdata('C:\Users\admin\Desktop\pso-lssvm\1_data\train_x.mat');
train_y=importdata('C:\Users\admin\Desktop\pso-lssvm\1_data\train_ctly.mat');
end
adapt.m代码(定义适应度函数):
function ad=adapt(gam,sig2,train_x,train_y)
ad=rclssvm(gam,sig2,train_x,train_y);
end
rclssvm.m代码(计算给定参数下模型拟合的均方误差):
function error=rclssvm(gam,sig2,train_x,train_y)
model= trainlssvm({train_x,train_y,'f',gam,sig2,'RBF_kernel'});
y=simlssvm(model,train_x);
error=sqrt(sum((train_y-y).^2)/length(train_y));
%error=sqrt(sum((train_y-y).^2)/numel(train_y));
end
程序运行后,每个粒子的最优位置和均方误差值存储在Pbest矩阵中,全局最优位置以及均方误差存储于gbest矩阵中。
做出对训练集的拟合效果图:
gam=300;%经过PSO优化后的两参数
sig2=49.5921519648569;
plotlssvm({train_x,train_y,'f',gam,sig2,'RBF_kernel'})
拟合图:
PSO算法介绍:http://blog.csdn.net/myarrow/article/details/51507671