what is reinforcement learning
Reinforcement learning is a way of training an AI system by letting it take actions in an environment and learn from the resulting rewards or penalties — rather than learning from a fixed set of labeled examples, it improves through repeated trial, feedback, and adjustment, similar to learning through practice and consequence.
```flowchart
{
"rows": [
{ "type": "nodes", "items": [
{ "title": "Agent takes an action in an environment", "sub": "no labeled example telling it the right answer", "tone": "blue" }
]},
{ "type": "arrow" },
{ "type": "group", "title": "Environment returns a reward or penalty", "items": [
{ "title": "Feedback signals whether that action was good", "sub": "not a direct correction, just an outcome signal", "tone": "orange" }
]},
{ "type": "arrow" },
{ "type": "nodes", "items": [
{ "title": "Agent adjusts to favor higher-reward actions", "sub": "improves through repeated trial and feedback", "tone": "green" }
]}
]
}
```
**Reinforcement learning is fundamentally different from other common AI training approaches because it learns from the consequences of actions rather than from labeled correct answers.** Many AI training methods rely on a large set of examples explicitly labeled with the correct output; reinforcement learning instead has an agent take actions within an environment and receive a reward or penalty signal reflecting how good that action turned out to be, gradually learning through repeated trial which actions tend to lead to better outcomes — a learning process much closer to how a person or animal might learn a skill through practice and consequence.
```svg
```
```svg
```
| Aspect | Supervised learning | Reinforcement learning |
|---|---|---|
| Learns from | Labeled correct examples | Rewards and penalties from actions taken |
| Feedback type | Direct correction | Outcome signal, not a direct answer |
| Learning process | Pattern matching against examples | Trial, feedback, and adjustment over time |
| Common use cases | Classification, prediction | Game playing, robotics, sequential decision-making |
**Reinforcement learning has become an important part of fine-tuning large language models, not just a technique for games and robotics.** Beyond its well-known use in training systems to play games or control robots, reinforcement learning is now widely used to fine-tune large language models based on human feedback about which responses are more helpful or appropriate — this application has become one of the most significant recent uses of reinforcement learning outside its traditional domains.
**Balancing exploration of new actions against exploitation of already-known good actions is one of reinforcement learning's central and persistent challenges.** An agent that only ever repeats actions it already knows work well risks missing a better action it hasn't tried, while an agent that experiments too much risks performing poorly by not using what it has already learned — this exploration-versus-exploitation tradeoff is a foundational challenge that reinforcement learning research continues to actively address.
**Reinforcement learning often requires substantially more training interactions than other AI training approaches, since it must learn from indirect outcome feedback rather than direct labeled answers.** Because reinforcement learning has to discover good actions through trial and error rather than being told the correct action directly, it frequently needs many more training interactions to learn effectively compared to approaches with direct labeled feedback — this data and computation intensity is a meaningful practical consideration when deciding whether reinforcement learning is the right approach for a given problem.
Read reinforcement learning through a practice-and-consequence lens: rather than being told the right answer directly, the system learns by trying actions and observing whether they lead to good or bad outcomes, gradually shaping its behavior toward whatever tends to earn the most reward over time.