%思路参考 http://blog.csdn.net/hujingshuang/article/details/46274723
clc
clear all
close all
load('Q2x');%读取训练数据
load('Q2y');%读取训练数据
q2x = Q2x;
q2y = Q2y;
weight=[];
q2x=[ones(size(q2x,1),1) q2x];%加一列偏置项
[m n]=size(q2x);
theta=zeros(n,m);
figure;
hold on;
plot(q2x(:,2),q2y,'.b');
l = q2x;%landmark 高斯核函数的标记
%在这里得到m*m的相似度函数
for i = 1:m
for j = 1:m
weight(i,j) = exp(-norm((q2x(i,:)-l(j,:)),2).^2/(2*3));%计算出第i个样本与第j个marker的相似度(类似SVM的高斯核),所以一行是同一个样本,一列是同一个TEST
end
end
options = optimset('MaxIter', 400);
initial_theta = zeros(n,1);
for i = 1:m
CostFunction = @(t)TrainCost(q2x,Q2y,t,weight(:,i));%每次都丢一组权值(同一个TEST)进去来得到一组参数Θ
theta(:,i)=fmincg(CostFunction, initial_theta, options);%返回最小的梯度THETA值一共有m列,得到m个样本拟合出来的Θ
end
y_fit=q2x*theta;
y_fit=diag(y_fit);
q2xs=sort(q2x);
result=[Q2x y_fit];
x_result=sortrows(result,1);
plot(x_result(:,1),x_result(:,2));
function [J grad] = TrainCost(X,y,Theta,weight)
m = size(X,1);
h = X*Theta;
J = (1/(2*m))*weight'*(h-y).^2;
grad = (1/m)*(weight.*(h-y))'*X;
grad = [grad(:,1);grad(:,2)];
end