线性回归的梯度下降不起作用

Gradient descent for linear regression not working

本文关键字:不起作用 回归 线性      更新时间:2023-10-16

我正在尝试为线性回归实现一个简单的梯度下降算法。我正在使用Armadillo C++线性代数库,我也是Armadillo的新手。这就是我要做的:

void linRegression(mat &features, mat &targets, double alpha,double error){
    mat theta = ones(features.n_cols+1);
    mat temp = zeros(features.n_cols+1);
    mat features_new = join_horiz(ones(features.n_rows),features);
    mat predictions;
    double con = alpha*(1.0/features.n_rows);
    int j = 0;
    while(j<1000){
        mat step_error = (features_new*theta - targets);
        for(unsigned int i=0;i<theta.n_rows;i++){
            temp(i) = con*sum(step_error%features_new.col(i));
        }
        theta = theta-temp;
        mat pred = predict(theta,features_new);
        cout<<theta<<endl;
        j++;
    }
}

但是θ的值一直在增加,最终达到无穷大。我不确定我做错了什么。

我认为while循环中的计算不正确。至少你可以在没有for循环的情况下做得更漂亮。以下是针对1个功能问题的简短代码:

#include <iostream>
#include <armadillo>
using namespace std;
using namespace arma;
int main(int argc, char** argv)
{
    mat features(10, 1);
    features << 6.110100 << endr
         << 5.527700 << endr
         << 8.518600 << endr
         << 7.003200 << endr
         << 5.859800 << endr
         << 8.382900 << endr
         << 7.476400 << endr
         << 8.578100 << endr
         << 6.486200 << endr
         << 5.054600 << endr;
    mat targets(10, 1);
    targets << 17.59200 << endr
        << 9.130200 << endr
        << 13.66200 << endr
        << 11.85400 << endr
        << 6.823300 << endr
        << 11.88600 << endr
        << 4.348300 << endr
        << 12.00000 << endr
        << 6.598700 << endr
        << 3.816600 << endr;
    mat theta = ones(features.n_cols + 1);
    mat features_new = join_horiz(ones(features.n_rows), features);
    double alpha = 0.01;
    double con = alpha*(1.0 / features.n_rows);
    int j = 0;
    while (j < 20000){
        mat step_error = (features_new*theta - targets);
        theta = theta - con * (features_new.t() * step_error);
        j++;
    }
    theta.print("theta:");
    system("pause");
    return 0;
}

程序返回:

theta:
   0.5083
   1.3425

通过正态方程法得到的结果是:

theta:
   0.5071
   1.3427

编辑

您的代码确实正确!问题可能出在特征归一化上。我将示例扩展到2个特征回归,并添加了规范化。如果没有规范化,它对我来说也不起作用。

#include <iostream>
#include <armadillo>
using namespace std;
using namespace arma;
int main(int argc, char** argv)
{
    mat features(10, 2);
    features << 2104 << 3 << endr
         << 1600 << 3 << endr
         << 2400 << 3 << endr
         << 1416 << 2 << endr
         << 3000 << 4 << endr
         << 1985 << 4 << endr
         << 1534 << 3 << endr
         << 1427 << 3 << endr
         << 1380 << 3 << endr
         << 1494 << 3 << endr;
    mat m = mean(features, 0);
    mat s = stddev(features, 0, 0);
    int i,  j;
    //normalization
    for (i = 0; i < features.n_rows; i++)
    {
        for (j = 0; j < features.n_cols; j++)
        {
            features(i, j) = (features(i, j) - m(j))/s(j);
        }
    }
    mat targets(10, 1);
    targets << 399900 << endr
        << 329900 << endr
        << 369000 << endr
        << 232000 << endr
        << 539900 << endr
        << 299900 << endr
        << 314900 << endr
        << 198999 << endr
        << 212000 << endr
        << 242500 << endr;

    mat theta = ones(features.n_cols + 1);
    mat features_new = join_horiz(ones(features.n_rows), features);
    double alpha = 0.01;
    double con = alpha*(1.0 / features.n_rows);
    while (j < 20000){
        mat step_error = (features_new*theta - targets);
        theta = theta - con * (features_new.t() * step_error);
        j++;
    }
    cout << theta << endl;
    system("pause");
    return 0;
}

结果:

θ:

  3.1390e+005
  9.9704e+004
 -5.6835e+003