Reinforcement learning concept

What is reinforcement learning?

Reference

What is reinforcement learning? It is a way for a learner, called an agent, to learn which actions help it reach a goal by interacting with an environment and using rewards as feedback. The agent aims to gain reward across a sequence of actions, so a choice can matter for its immediate result and for the situations it encounters later.

Reviewed 2026-10-09

A learner improves through feedback

Reinforcement learning is a kind of machine learning in which an agent learns what to do by interacting with an environment. The agent is not given the correct action for every situation; it tries actions and uses feedback to improve its choices. That feedback evaluates behavior, rather than telling the agent which action was correct. The approach is useful for describing learning problems where the learner must act and receive consequences as it goes.

The feedback is a reward signal, a number provided by the environment. The aim is to gain cumulative reward, also called return. A reward describes how the task values an outcome or action; by itself, it does not tell the agent exactly what to do. The agent has to learn which choices tend to lead to reward. The reward goal is part of how the task is set up, so it matters what the reward captures.

Actions lead to new situations

At each step, the agent gets an observation of the environment, chooses an action according to a policy (a rule for choosing actions), and receives a new observation and reward. The action may change the environment, which means the next choice happens in a situation shaped by earlier choices. The sequence of observations, actions and rewards gives the learner experience to use. In some problems, an observation is only a partial view of the environment.

The agent learns from these sequences of actions and feedback. It may explore by trying actions to learn about their results, then use what it has learned to pursue reward. Since rewards can come later, a choice that earns less now may still help produce more reward over time. This is why the goal is usually described across a trajectory, or sequence of steps, rather than as choosing the action with the biggest immediate payoff.

A different kind of machine learning feedback

Reinforcement learning is one approach within machine learning. In supervised learning, training examples provide target answers, such as the correct label or action. In reinforcement learning, feedback evaluates actions the agent took rather than specifying the correct action in advance. The difference is about the learning signal: one supplies target answers, while the other supplies evaluative feedback during interaction.

Reinforcement learning is useful to think about when actions affect what happens next and the goal is defined in terms of reward over time. It sits alongside other machine learning approaches, but it describes a different setup for learning. A task may involve a sequence of decisions, with each decision affecting later situations or available rewards. For more on the parent subject, see reinforcement learning.

Questions people ask

Does reinforcement learning tell the agent the correct action?

Usually, no. Its feedback evaluates an action the agent took; it does not directly specify the correct action for that situation.

Why does reinforcement learning consider future rewards?

An action may change the next situation and affect rewards later. The goal is cumulative reward across a sequence, not just the immediate reward.

How is reinforcement learning different from supervised learning?

Supervised learning uses examples with target answers. Reinforcement learning uses feedback about actions taken and aims to maximize reward through interaction.

Learn more

Part 1: Key Concepts in RL
A concise introduction to the agent-environment loop, reward, and cumulative return.
5.1 Reinforcement Learning
An explanation of exploration feedback and how an agent uses it to estimate a policy.

What changed

Sources

Part of Reinforcement learning: The parts of a reward-based problem Actions have consequences over time Learn what works, then evaluate the tradeoffs