使用梯度下降算法时的初值误差

Question

问题： 初始值为 10000，解收敛到 10000 而不是实际解 1。

import numpy.linalg as nl
x_ini=10000

def obj(x):
    f = x**2 - 2*x + 3
    return f

def grad(x):
    df = 2*x - 2
    return df

n_iter=1000
lr=0.001

x_old = x_ini


for _ in range(True):
    
    x_new = x_old - lr*(grad(x_old))
    x_old = x_new
    
    if(nl.norm(grad(x_old))<=1e-03):
        break
    x_new = x_old
    
print(x_new)

Answer 1

while True:
    
    x_new = x_old - lr*(grad(x_old))
    x_old = x_new
    y = nl.norm(grad(x_old))
    if(y<=1e-03):
        break
    x_new = x_old

print(x_new)

您可以将 for 更改为 while