Reinforcement learning

Reinforcement learning is a machine learning approach in which an agent learns a strategy, called a policy, by taking actions in an environment and receiving rewards or penalties as feedback, then adjusting its behavior to earn more reward over time. Unlike supervised learning, there is no dataset of correct answers to imitate; the agent must discover, through trial and error, which actions lead to good outcomes, often after a delay.

A core tension in reinforcement learning is exploration versus exploitation: trying new, uncertain actions that might reveal a better strategy, exploration, versus repeating actions already known to work well, exploitation. A simpler, related technique, the multi-armed bandit, handles this trade-off for problems with no meaningful state, such as choosing which ad to show next, while full reinforcement learning also accounts for how actions change the state of the environment over time.

Reinforcement learning has driven notable results in games and robotics, and is used in analytics-adjacent optimization modeling problems such as dynamic pricing, recommendation ranking and resource allocation, where decisions unfold sequentially and each action affects future options. It typically requires either a realistic simulated environment or a large amount of real-world trial and error, both of which can be costly, slow or risky, and reward functions that fail to capture what actually matters can lead the agent to find unintended shortcuts.