Reinforcement learning (RL) is a branch of machine learning where an agent learns to make decisions by interacting with an environment. After each action, the agent receives a numerical reward (or penalty) and updates its behavior to favor actions that lead to better long-term outcomes. Unlike supervised learning, the agent is not given labeled examples of correct answers — it must discover effective strategies through trial and error.
How Reinforcement Learning works
At each step, the agent observes the current state of the environment, chooses an action from its available options, and then receives a reward along with the next state. The goal is to learn a policy, essentially a mapping from states to actions, that maximizes the expected sum of future rewards. Techniques such as Q-learning estimate the value of taking each action in each state, while policy-gradient methods directly adjust the policy based on which actions tend to produce high rewards. Modern approaches combine RL with deep neural networks — as in Deep Q-Networks (DQN) — to handle problems with very large or continuous state spaces, such as raw video input.
Why it matters
Reinforcement learning powers many of the most visible breakthroughs in AI, from game-playing systems like AlphaGo and AlphaZero to the fine-tuning step behind modern large language model assistants through methods such as RLHF (Reinforcement Learning from Human Feedback). It is also used in robotics, autonomous driving, recommendation systems, supply-chain optimization, and resource scheduling, anywhere a system must make a sequence of decisions whose effects unfold over time and where the best long-term strategy is not obvious in advance.
Key types
- Model-free RL: the agent learns directly from experience without building an internal model of the environment (e.g., Q-learning, PPO).
- Model-based RL: the agent learns a model of how the environment works and plans actions using that model.
- Policy-gradient methods: directly optimize the policy, useful for continuous actions and stochastic policies.
- Multi-agent RL: several agents learn simultaneously in a shared environment, useful for game theory and coordination.
Reinforcement learning remains one of the most flexible frameworks for sequential decision-making, and it is increasingly the bridge between pattern-recognition models and systems that act autonomously in the real world. The canonical reference text is Sutton and Barto's "Reinforcement Learning: An Introduction".