Definition
Reinforcement learning trains an agent to maximize cumulative reward through interaction with an environment. Unlike supervised learning, there is no fixed correct label for every state — only feedback after actions.
RL underpins game-playing AIs, robotics control, and alignment methods like RLHF for language models.
In simple terms
RL is like training a dog with treats: good actions get rewards, bad ones get none (or penalties). Over many trials, the dog learns which behaviors pay off.
Where you see it
- AlphaGo and game agents learning through self-play.
- Robotics policies for grasping and locomotion.
- RLHF steering chat models toward preferred answers.
How it works
1.Observe state
The agent sees the current environment.
2.Take action
Choose based on a policy (strategy).
3.Receive reward
Numeric feedback signals success or failure.
4.Update policy
Adjust toward actions that increase long-term reward.
Why it matters
- RL is how AI systems learn sequential decisions when labels are scarce but outcomes can be scored.
Often confused
RL is the same as supervised fine-tuning.
Supervised learning copies labeled examples; RL optimizes a reward signal through exploration.