reinforcement learning (RL)

basicwidespread

A machine learning method where the program discovers good behavior through trial and error based on rewards and penalties.

In plain terms

Think of how you teach a dog to sit: you do not explain the movement to it, but give it a treat when it gets it right. Reinforcement learning does the same with machines: the program tries, gains points for good moves, loses points for bad ones, and over time learns what leads to success.

It is useful when we are not looking for a single good answer, but must make a long series of decisions well: winning a game, controlling a robotic arm, or efficiently managing a building's cooling. In such cases, no one can predict in advance whether each step is correct; often only the final result is visible.

This differs from other types of machine learning. Most models learn from prepared examples where humans marked the correct answer. Here there is no answer key: the system learns from its own experiences and sometimes finds solutions that never occurred to humans.

How it works

There is a learning program, the agent, and an environment in which it moves: a game, a simulation, or a part of the real world.
The agent observes the situation and chooses a step, initially essentially at random.
The environment provides feedback: the situation changes, and the agent receives a reward or penalty, meaning a number indicating how good the step was.
Based on this, the agent adjusts its rules: it chooses more frequently what was followed by reward, and less frequently what was followed by punishment. Meanwhile, it must balance repeating proven steps with trying new things.
This loop repeats thousands or millions of times, mostly in simulation, until a strategy emerges that yields the most reward in the long run.

An everyday example

When DeepMind began optimizing the cooling of Google data centers in 2016, the system tried settings for fans and cooling units, and received a reward if the room stayed cool enough with less energy. This significantly reduced the energy consumption of cooling. Similarly, the movement of more and more robotic arms is being trained in simulation.

Who is affected

everyoneFine-tuning chatbots usually relies on this method, and some news feeds and video recommenders also use it to learn from your reactions.

industrial worker, engineerRobot motion, production line control, and energy system management are increasingly optimized this way. → Industry

developerOne important step in training language models and coding agents is based on reinforcement learning, so it is worth understanding how it works and its pitfalls. → Code

entrepreneur, executiveIn pricing, inventory management, and recommendation systems, the system will precisely optimize for what you reward it with, making goal setting a managerial decision.

A common misconception

MythThe machine feels reward and punishment just like a human or an animal.

In realityThe reward here is just a number that the program tries to maximize. It contains no joy or pain; the terminology is a metaphor borrowed from psychology.

Limits and risks

It requires countless attempts, so it is mostly trained in simulation, and what works there does not always work in reality.
The system maximizes what you measure, not what you actually want: with a poorly defined reward, it finds loopholes and chases the score instead of the goal.
In the real world, trial and error can be expensive or dangerous, so it is only applicable under strict limits in areas such as driving, healthcare, or finance.
It is difficult to explain why the system made a particular decision, and it may behave unpredictably in new situations not seen during training.

Similar, but not the same

supervised learning – There, a human provides the correct answer for each example. Here, no correct answer is given; only a reward indicates how good a single step or the entire sequence of steps was.
unsupervised learning – It searches for patterns in data without feedback, whereas reinforcement learning acts and learns from the consequences of its actions.
fine-tuning – Fine-tuning refers to further training a pre-trained model by any method. Reinforcement learning is an independent learning method that can also be a tool for fine-tuning.

Related concepts

Broader

machine learning, artificial intelligence

Narrower

Deep reinforcement learning, Q-learning, reinforcement learning from human feedback, multi-armed bandit

On the edge

supervised learning, unsupervised learning, neural network, AI-agent, reward function

A brief history

The idea grew out of the psychology of animal learning and control theory. The field's foundational textbook by Richard Sutton and Andrew Barto was published in 1998. The general public first took notice in 2016 when DeepMind's AlphaGo program defeated Lee Sedol, one of the best Go players from South Korea. From 2024, reinforcement learning has become a central tool for training reasoning language models. Sutton and Barto received the 2024 Turing Award for laying the foundation of the field.

Where things stand

By the end of 2026, it is an established method in games, recommendation systems, robotics, and the training of large language models, especially for teaching reasoning models and agents. In the physical world, such as in autonomous driving and industrial control, it is spreading more cautiously, mainly with pre-training in simulation.

Status: 2026 November

Also known as

reinforcement learning · reinforcement learning · Reinforcement learning · reinforcement learning · RL

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.