Sensory input → Reward or punishment

RL Framework:

It can be thought of a cost function.

Not supervised learning:

Not unsupervised learning:

Trial and error is essential for adapting the parameters to find the gradient.

Temporal difference:

Prediction made at t+1 is probably more accurate than the prediction made at t. So go forward in time and learn what to do at t.

Tic tac toe example:

X is the RL, O is the opponents.