PSO-LSSVM的matlab实现

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 ad<Pbest(i,3)%更新当前微粒的最好位置
            Pbest(i,1)=gam;
            Pbest(i,2)=sig2;
            Pbest(i,3)=ad;
        end
        if Pbest(i,3)<gbest(1,3)%更新全局的最好位置
            gbest(1,1)=Pbest(i,1);
            gbest(1,2)=Pbest(i,2);
            gbest(1,3)=Pbest(i,3);
        end
        %调整微粒速度与位置,同时限定着参数范围
        W=W_min+(W_max-W_min)*(1-j/K);
        P_v(i,1)=W*P_v(i,1)+C1*rand*(Pbest(i,1)-P_xy(i,1))+C2*rand*(gbest(1,1)-P_xy(i,1));
        if(P_v(i,1)>V_max)
            P_v(i,1)=V_max;
        end
        if(P_v(i,1)<V_min)
            P_v(i,1)=V_min;
        end
        P_v(i,2)=W*P_v(i,2)+C1*rand*(Pbest(i,2)-P_xy(i,2))+C2*rand*(gbest(1,2)-P_xy(i,2));
        if(P_v(i,2)>V_max)
            P_v(i,2)=V_max;
        end
        if(P_v(i,2)<V_min)
            P_v(i,2)=V_min;
        end
        P_xy(i,1)=P_xy(i,1)+P_v(i,1);
        if(P_xy(i,1)>gammax)
            P_xy(i,1)=gammax;
        end
        if(P_xy(i,1)<gammin)
            P_xy(i,1)=gammin;
        end
        P_xy(i,2)=P_xy(i,2)+P_v(i,2);
        if(P_xy(i,2)>sig2max)
            P_xy(i,2)=sig2max;
        end
        if(P_xy(i,2)<sig2min)
            P_xy(i,2)=sig2min;
        end
    end
end

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