Reinforcement Learning

Reinforcement learning is a branch of machine learning where a program, called an agent, learns to make good decisions by acting in an environment and receiving rewards or penalties for the outcomes, instead of being shown correct answers in advance.

itArtificial intelligence and machine learning

Don't Panic — Reinforcement Learning

Reinforcement learning is a way to train a decision maker by letting it act in the world, watching what happens, and adjusting. No labeled examples, no teacher pointing at the right answer. Just an agent, an environment, and a reward score that tells it whether things went well.

The problem it solves is sequential decision-making: situations where each choice reshapes the situations that follow, and where the right answer depends on what you did three steps ago. Before reinforcement learning, people solved these problems with hand-coded rules or brute-force search. Rules are brittle; search is expensive. An agent that learns from its own experience sidesteps both.

Everything else hangs off one loop. The agent observes a state, picks an action according to its policy, the environment returns a reward and a new state, and the agent updates. The Bellman equation says the value of any state is the immediate reward plus the discounted value of wherever you land next. That recursive idea, and the question of how to compute it when you cannot enumerate every state, is the entire intellectual content of the field.

The thing that will surprise you is how literal the agent is. It optimizes exactly the reward you give it, not the outcome you had in mind. A boat trained to finish a race can learn to circle the track collecting bonus blocks instead. The gap between what you rewarded and what you meant is where nearly every real-world failure lives. Designing the reward function is harder than choosing the algorithm.

If you want to know what the vocabulary means, the Cheatsheet tab has every term and formula in one place. If you want to know what teams get wrong when they actually try this, the Field Notes tab has the judgment. If you want to see the timeline of how this field went from a 1957 paper to beating the world champion at Go, that is the Timeline tab. And if you want to see what tools practitioners actually use, the Landscape tab covers the frameworks, environments, and platforms that make up the modern reinforcement learning stack.

Don't Panic — Reinforcement Learning

Relevant careers