What is the core idea behind reinforcement learning?
Reinforcement learning learns by trial and error, maximising reward over time rather than copying labelled examples.
How does reinforcement learning differ from related concepts?
| Concept | Difference |
|---|---|
| Reinforcement vs Supervised Learning | Supervised learning learns from labelled answers. Reinforcement learning learns from rewards for its own actions. |
| Reinforcement Learning vs RLHF | RLHF is reinforcement learning where the reward comes from human preferences, used to align language models. |
| Reinforcement vs Unsupervised Learning | Unsupervised learning finds structure in data. Reinforcement learning optimises behaviour toward a goal. |
How does reinforcement learning work?
- An agent observes the state of an environment
- It takes an action and receives a reward or penalty
- It updates its policy to favour actions that lead to more reward
- Over many episodes it learns a strategy that maximises long-term reward
Where is reinforcement learning used?
- Game-playing systems like AlphaGo
- Robotics and control
- Aligning large language models through RLHF
- Training reasoning models on verifiable tasks
Why is reinforcement learning important?
Reinforcement learning is how AI learns behaviour rather than just patterns. It aligned modern chatbots through RLHF, and in 2025 and 2026 it became central to training reasoning models that think step by step on math, code, and science.
How is reinforcement learning used in practice?
It powers game AI, robotics, recommendation, and the alignment and reasoning stages of large language models. Its challenges include sample inefficiency, unstable training, and the difficulty of designing rewards that capture what you actually want.
Frequently Asked Questions
How is reinforcement learning different from other machine learning?
Supervised and unsupervised learning learn from a fixed dataset. Reinforcement learning learns by interacting with an environment and optimising for reward.
Is RLHF the same as reinforcement learning?
RLHF is a specific application of reinforcement learning where human preferences provide the reward signal, used to make language models helpful and safe.