Reinforcement Learning
What is it?
Reinforcement learning trains an agent to make decisions by trial and error, rewarding good outcomes and penalizing bad ones until it learns a strategy.
Explain like I'm 5
Why was it created?
Some problems have no labeled 'right answer' per step — only a goal and feedback over time. RL was developed to learn good behavior from rewards rather than labels.
Where is it used?
- Game-playing AI
- Robotics and control
- Recommendation and bidding
- Aligning LLMs (RLHF)
Why should developers care?
RL underlies game-playing breakthroughs, robotics, and the RLHF step that aligns chat models, so it's key to how AI learns to act.
How does it work?
An agent observes a state, takes an action, and receives a reward and a new state. It learns a policy — a mapping from states to actions — that maximizes total reward over time, balancing exploration (trying new things) with exploitation (using what works).
Real-world example
An AI learns to play a game by playing millions of rounds, gradually favoring moves that led to winning.
Common use cases
- Sequential decision-making
- Control and robotics
- Optimization under feedback
- Preference alignment (RLHF)
Advantages
- Learns from goals, not labels
- Handles sequential decisions
- Can discover novel strategies
- Adapts to feedback over time
Disadvantages
- Sample-inefficient (needs many trials)
- Reward design is hard and gameable
- Training can be unstable
- Hard to test safely in the real world
When should you use it?
When the task is sequential decision-making with a reward signal rather than labeled correct answers.
When should you avoid it?
When you have labeled data for direct prediction (use supervised learning) or can't define a safe, meaningful reward.
Alternatives
Related terms
Interview questions
Beginner
- What does an RL agent learn from?
- What is a reward in RL?
Intermediate
- What is the exploration vs exploitation trade-off?
- What is a policy?
Senior
- Why is reward design so error-prone?
- How does RLHF apply RL to language models?
Common misconceptions
- "RL needs labeled answers" — it learns from rewards, not per-step labels.
- "A reward function is easy to write" — poorly designed rewards get gamed in surprising ways.
Fun facts
- RL agents mastered Go and Atari games at superhuman levels.
- The RL in RLHF is what turns a base LLM into a preference-aligned assistant.
Timeline
- 1980s-90s — Foundations of RL (Q-learning, policy methods)
- 2015-2016 — Deep RL masters Atari and Go
Learning resources
Quick summary
Reinforcement learning trains an agent by trial and error using rewards, learning a policy that maximizes long-term payoff.
Cheat sheet
- Learns from rewards, not labels
- State → action → reward loop
- Explore vs exploit
- Powers game AI and RLHF