Some of the most important ideas in computing begin with a simple question: how can a machine improve when nobody tells it the right answer every time?
For computer scientists Richard S. Sutton and Andrew G. Barto, the answer was reinforcement learning—a field that allows an artificial agent to learn by interacting with an environment, receiving feedback, and adjusting its behavior over time.
In 2025, the Association for Computing Machinery awarded Sutton and Barto the A.M. Turing Award, often described as computing’s highest honor, for developing the conceptual and algorithmic foundations of reinforcement learning. Their work helped establish a different path for machine learning: instead of learning only from labeled examples, a system could learn from the consequences of its choices.
Learning from consequences
The basic arrangement is familiar from everyday life. An agent takes an action, the world responds, and the agent receives a reward or penalty. A computer playing a game might gain points for making progress and lose points for making a mistake. A robot might be rewarded for reaching a target while avoiding obstacles.
The challenge is that the best choice is not always the one that produces an immediate reward. A move in a game may look unhelpful now but create an advantage several turns later. Reinforcement learning therefore tries to estimate which actions are likely to lead to better results over time.
Sutton’s work on temporal-difference learning gave computers a practical way to update their expectations as new information arrived. Rather than waiting until the end of an entire task, an algorithm could learn from the difference between what it expected and what happened next. Barto’s research helped connect these learning methods with ideas about control, decision-making, and how an agent can choose actions.
A different route to machine intelligence
The approach became especially influential in games, where rules are clear and progress can be measured. Reinforcement-learning systems have learned to play board games, video games, and other controlled environments by combining exploration—trying unfamiliar actions—with the use of strategies that already appear promising.
Google DeepMind’s AlphaGo offered a widely seen example. The system combined several machine-learning techniques, including reinforcement learning, to improve its play of the ancient board game Go. Its 2016 victory over champion Lee Sedol showed how systems could discover effective strategies that were not simply copied from a database of human moves.
Reinforcement learning has also influenced robotics, resource management, recommendation systems, and scientific research. But it is not a magic shortcut to general intelligence. Learning can require enormous amounts of practice, and a poorly designed reward can encourage a system to exploit the rules rather than achieve the goal people intended.
Why the idea still matters
The lasting contribution of Sutton and Barto is not one particular program. It is a general framework for thinking about learning as a continuing process of decision and feedback.
Their textbook, Reinforcement Learning: An Introduction, helped make that framework accessible to generations of students and researchers. Its central insight remains useful because many important problems are not solved by choosing one correct answer. They involve acting under uncertainty, learning from experience, and balancing immediate results against future possibilities.
That makes reinforcement learning one of computing’s most durable ideas: a bridge between prediction and action, built around the possibility that machines can become better not simply by being told what to do, but by learning what happens when they try.
