A computer program is said to learn from experience E with respect to some task T and some performance measure P, if its performance on T, as measured by P, improves with experience E.
- Experience E
- task T
- performance measure P
-
Supervised Learning
Notation
- m: the number of training examples
- x: input variable / features
- y: output variable / target variable
- (x, y) : one training example
- (x^(i), y^(i)) : the i training example
- h: hypothesis, map from x to y
different problem
-
Regression problem--predict real-valued outputs
- Linear Regression
-
Classification problem--predict discrete-valued outputs
- Logistic Regression
cost function(代价函数)--J
Gradient descent
-
Gradient descent alogrithms
Feature Scaling
Get every feature into approximately a [-1, +1] range.
Learning rate
- rate is too small: slow convergence
- rate is too large: cost function may not decrease on every iteration and may not converge.
-
Normal Equation Method
-
Unsupervised Learning
- No supervised, Only a reward signal
- feedback is delay
- Time really matters