Reinforcement Learning: Model-free

Simply put, reinforcement learning (RL) is a term used to indicate a large family of dierent algorithms RL that all share two key properties. First, the objective of RL is to learn appropriate behavior through trialand-error experience in a task. Second, in RL, the feedback available to the learning agent is restricted to a reward signal that indicates how well the agent is behaving, but does not indicate specifically how the reward signal agent could improve its behavior. For example, consider writing an essay for a course and receiving a numerical score in the range 0–100. If your score is less-than-perfect, you know that your performance could be improved upon, but the feedback itself doesn’t indicate specifically how your essay should have been dierent. In more complex cases, optimal behavior will generally require numerous separate decisions, and there may be delayed or missing reward signals. For example, one can imagine trying to teach a computer to play checkers by providing a positive reward signal every time it wins a game, and a negative (penalty) signal every time it loses. In this case, each individual action (moving a piece) is not rewarded, and the only reward signal is provided at the end of a game. Learning how to improve behavior given this limited type of feedback is both the goal and challenge facing all RL algorithms.