---
title: 'Reinforcement learning (RL)'
description: 'A family of methods in which an agent learns a policy by interacting with an environment and optimizing expected cumulative reward.'
canonical_url: 'https://darkfactory.dev/glossary/reinforcement-learning'
markdown_url: 'https://darkfactory.dev/glossary/reinforcement-learning.md'
collection: glossary
date_published: '2026-08-03T00:00:00-04:00'
date_modified: '2026-08-26T00:00:00-04:00'
---

# Reinforcement learning (RL)


## Definition

Reinforcement learning is a family of methods in which an agent learns how to act through consequences. At each step, the agent receives a state or observation, chooses an action, and receives a reward plus a new state. A policy maps observations or states to actions. Training seeks a policy that maximizes expected cumulative reward, often called return, rather than the immediate reward from one step.

This sequential setup creates two central problems. Credit assignment asks which earlier actions deserve responsibility for a later result. Exploration asks when to try uncertain actions that may teach the agent something rather than repeat the action that currently looks best. Value functions, policy gradients, temporal-difference methods, and model-based planning address different parts of this problem.

Sutton and Barto's textbook provides the standard modern vocabulary and traces the field through trial-and-error learning, optimal control, and related research. It does not present reinforcement learning as one person's invention, and the historical record does not justify a simple coinage claim.

## Distinguish it from nearby terms

- **Supervised learning** fits input-output examples supplied as targets. Reinforcement learning learns from reward and state transitions, which may be delayed and affected by the agent's own actions.
- A **bandit** problem chooses among actions without modeling a longer state sequence. It is a simpler relative of full reinforcement learning.
- **Planning** can use a known or learned model of the environment without updating a policy from live experience.
- **Reinforcement learning from human feedback** uses human preferences or judgments within a larger training pipeline. Human feedback alone does not make every method reinforcement learning.

## Operational significance

The reward is a specification. If it omits a real constraint, the learned policy may optimize the score while harming the intended outcome. Training can also be costly or unsafe when exploration affects people or production systems. Simulation, offline data, constrained action spaces, monitoring, and stop conditions are part of the design, not deployment details to add later.

## Check your understanding

An LLM agent calls tools and receives results during a workflow. If its policy or parameters do not update from those results, the run is agentic behavior, not reinforcement learning.

## Also called

RL

## Related terms

- [Reward hacking](https://darkfactory.dev/glossary/reward-hacking)
- [Reinforcement learning from human feedback (RLHF)](https://darkfactory.dev/glossary/rlhf)

## Evidence and further reading

- [NIST AI 100-2: Adversarial Machine Learning](https://csrc.nist.gov/pubs/ai/100/2/e2025/final)
- [Google Machine Learning Glossary](https://developers.google.com/machine-learning/glossary/)
- [Reinforcement Learning: An Introduction](https://mitpress.mit.edu/9780262039246/reinforcement-learning/)