Reinforcement Learning
Reinforcement learning is learning by trial, error and reward. A program tries actions, sees what happens, and slowly learns which actions give the most reward. It trains game players, robots and even helps to train chatbots.
New words
- Agent
- The learner that chooses actions, like a player in a game.
- Environment
- Everything the agent acts in. It gives back a new state and a reward.
- Reward
- A number that tells the agent how good its last action was.
- Policy
- The rule the agent follows to choose an action in each state.
- Value
- How much total reward the agent expects to get from a state.
- Q-value
- The expected total reward of taking one action in one state.
- Explore and exploit
- Try something new, or use the best option you know.
What you will learn
- Learning without a teacher
- The five parts of every RL problem
- The goal is total reward, not the next reward
- Policy and value
- Explore or exploit?
- Q-learning in simple words
- Try it: a grid world
- Where RL is used, and why it is hard
- Trade-offs
- Interview tip
- Key takeaways
10 quiz questions at the end.
The full lesson, with interactive demos and the quiz, opens after you sign in with Google. It is free.