常见优化算法实现
一、梯度下降展开的是loss^2的导数
不知道这个函数: y = exp(1*x^{2}+2*x+1) 为什么会不能优化。。。
'''
Description: 梯度下降法优化非线性函数
Author: suyunzheng
Date: 2021-10-06 23:06:34
LastEditTime: 2021-10-06 23:38:21
LastEditors: maple
'''
import os
import sys
import numpy as np
import math
import random
def generator_sample(a, b, c):
x_list =
# print(x_list)
y_list =
print(y_list)
return x_list, y_list
def GD_optimizer(x_list, y_list, epoch, a, b, c, lr):
print("a:{}, b:{}, c:{}".format(a, b, c))
for i in range(epoch):
J_T =
loss = 0
# pass
for (x_i, y_i) in zip(x_list, y_list):
y_hat_i = math.exp(a*x_i*x_i+b*x_i+c)
# print("y_i:{}, y_hat_i:{}".format(y_i, y_hat_i))
loss_i = y_i - y_hat_i
loss+=loss_i
# print("loss_:{}".format(loss_i))
commom_ele = -2*loss_i*y_hat_i
J_T+=commom_ele*x_i*x_i
J_T+=commom_ele*x_i
J_T+=commom_ele
delta_a = 1*lr*J_T
delta_b = 1*lr*J_T
delta_c = 1*lr*J_T
# 更新a b c
a+=delta_a
b+=delta_b
c+=delta_c
print("Iter: {}, a: {}, b:{}, c:{}, loss: {}".format(i+1, a, b, c, loss))
pass
if __name__ == '__main__':
x_list, y_list = generator_sample(1,2,1)
a = 1
b = 1
c = 1
GD_optimizer(x_list, y_list, 100, a, b, c, 0.1)下面对 y=x^{2}+1 优化:
'''
Description: 梯度下降法优化非线性函数
Author: suyunzheng
Date: 2021-10-06 23:06:34
LastEditTime: 2021-10-07 14:57:21
LastEditors: maple
'''
import os
import sys
import numpy as np
import math
import random
import matplotlib.pyplot as plt
def show_x_y(x_list, y_list):
plt.scatter(x_list, y_list, s = 0.5)
plt.show()
def generator_sample(a, b, c):
x_list =
# print(x_list)
# y_list =
y_list =
# print(y_list)
show_x_y(x_list, y_list)
return x_list, y_list
def GD_optimizer(x_list, y_list, epoch, a, b, c, lr):
print("a:{}, b:{}, c:{}".format(a, b, c))
for i in range(epoch):
J_T =
loss = 0
# pass
for (x_i, y_i) in zip(x_list, y_list):
# y_hat_i = math.exp(a*x_i*x_i+b*x_i+c)
y_hat_i = a*x_i*x_i+0*x_i+c
# print("y_i:{}, y_hat_i:{}".format(y_i, y_hat_i))
# pass
loss_i = math.pow((y_hat_i - y_i),2)
loss+=loss_i
# print("loss_:{}".format(loss_i))
# commom_ele = 2*(y_hat_i - y_i)*y_hat_i
commom_ele = 2*(y_hat_i - y_i)
# commom_ele = y_hat_i
J_T+=x_i*x_i*commom_ele
# J_T+=x_i*commom_ele
J_T+=1*commom_ele
delta_a = -1*lr*J_T
# delta_b = -1*lr*J_T
delta_c = -1*lr*J_T
# 更新a b c
a+=delta_a
# b+=delta_b
c+=delta_c
print("Iter: {}, a: {}, b:{}, c:{}, lr:{}, loss: {}".format(i+1, a, b, c, lr, loss))
pass
if __name__ == '__main__':
x_list, y_list = generator_sample(1,0,1)
a = 0.1
b = 0.0
c = 0.3
GD_optimizer(x_list, y_list, 1000, a, b, c, 0.0001)二、牛顿法
三、高斯牛顿法
参考:
/*
* @Description: 常见优化算法的实现
* @Author: suyunzheng
* @Date: 2021-10-05 21:50:27
* @LastEditTime: 2021-10-05 22:55:12
* @LastEditors: maple
*/
#include <iostream>
#include <vector>
#include <eigen3/Eigen/Core>
#include <eigen3/Eigen/Dense>
using namespace std;
constexpr int N = 100;
void generator_sample(const vector<int>& params, vector<double>& x_data, vector<double>& y_data){
int a = params;
int b = params;
int c = params;
for(int i = 0; i<N; i++){
double x = i/100.0;
x_data.push_back(x);
y_data.push_back(exp(a*x*x+b*x+c)+(rand()%100)/100);
}
for(int i = 0; i<N; i++){
cout << x_data << &#34;, &#34; << y_data << endl;
}
cout << endl;
}
int GN(vector<double>& parms, vector<double>& x_data, vector<double>& y_data){
int iter_num = 100;
double& a = parms;
double& b = parms;
double& c = parms;
double pre_cost = 0;
double cost = INT_MAX;
for(int i = 0; i<iter_num; i++){
double eps = pre_cost - cost;
cout << &#34;Iter: &#34; << i <<&#34;, cost: &#34; << cost << &#34;, eps: &#34; << eps << endl;
if(cost<=1e-10 || abs(eps)<=1e-10 && i>=30){
cout << &#34;convergence&#34; << endl;
break;
}
pre_cost = cost;
cost = 0; // cost清零
Eigen::Matrix3d H_ = Eigen::Matrix3d::Zero();
Eigen::Vector3d b_ = Eigen::Vector3d::Zero();
for(int i = 0; i<x_data.size(); i++){
Eigen::Matrix3d tmp_H = Eigen::Matrix3d::Zero();
Eigen::Vector3d tmp_b = Eigen::Vector3d::Zero();
Eigen::Vector3d J_T = Eigen::Vector3d::Zero();
double x = x_data;
double y = y_data;
double exp_ele = exp(a*x*x+ b*x + c );
J_T.x() = -exp_ele*x*x;
J_T.y() = -exp_ele*x;
J_T.z() = -exp_ele;
// H和b求和
H_+=J_T*J_T.transpose();;
b_+=-1*J_T*(y-exp_ele);
cost+=pow(y-exp_ele, 2);
}
// cout << H_ << endl;
// Eigen::Vector3d delta = H_.inverse()*b_;
// Eigen::Vector3d delta = H_*b_;
Eigen::Vector3d delta = H_.ldlt().solve(b_);
a+=delta.x();
b+=delta.y();
c+=delta.z();
// cout << &#34;a: &#34; << a << &#34;, b: &#34; << b << &#34;, c: &#34; << c << endl;
}
cout << &#34;a: &#34; << a << &#34;, b: &#34; << b << &#34;, c: &#34; << c << endl;
}
int main(){
vector<int> params = {1,2,1};
vector<double> x_data;
vector<double> y_data;
generator_sample(params, x_data, y_data);
vector<double> guess_params = {0, 0 , 0};
// 优化
GN(guess_params, x_data, y_data);
return 0;
}
四、L-M算法
页:
[1]