regularization
通过让各\(\theta_i(i>0)\)值尽量小一些, 可以避免overfitting (why?)
在\(J(\theta)\)上增加\(\frac{\lambda}{2m} \sum_{i=1}^n \theta_i^2\) (注意\(\theta_0\)不需要加)
注意这里分母\(2m\)不是\(2n\) (why?)
但如果\(\lambda\)过大,则会导致underfitting(\(\theta\)都趋于0了)
随着\(\lambda\) 增大,train err单调增大 (why?)
对于梯度下降法:
这样的话\(gradient(i)\)就多了\(\alpha \frac {\lambda} m \theta\) (注意\(gradient(0)\)不变)
调整后J仍然bowl shape
function [J, grad] = costFunctionReg(theta, X, Y, lambda)
m = length(Y);
n = length(theta);
H = sigmoid(X * theta);
J = -1 / m * (Y' * log(H) + (1.-Y)' * log(1.-H)) + lambda / (2*m) * sum(theta(2:n).^2);
grad = 1 / m * (X' * (H - Y) + [0; (lambda .* theta(2:n))]);
end
对于normal equation:
\(\left(X^T X+\lambda \begin{bmatrix} 0\\&1\\&&\ddots\\&&&1\end{bmatrix}\right)^{-1} X^T Y\) (why?)