Reinforcement Learning

Reinforcement Learning trains an agent through actions and numeric rewards rather than a labeled dataset.

The agent takes an action. The environment returns a new state and a reward or penalty. Over many iterations the agent builds a policy: which action to take in a given state to maximize long-term cumulative reward. The only objective is total reward over time. Trial and error can find strategies nobody wrote down.

RL powered DeepMind's AlphaGo, which beat the world champion at Go with lines of play humans had not catalogued in thousands of years of the game. The same family of methods trains robots to walk, cross obstacle courses, and drive through messy traffic.

It is also the learning step inside RLHF, which makes large language models more useful and less harmful by optimizing against preference-derived rewards.

State, action, reward, next state: that tuple is the dataset RL creates by acting. There is no need for a human to label the correct action in every state. The policy maps states to actions to maximize discounted return. AlphaGo’s value and policy nets, plus tree search, used that loop to beat the world champion at Go.

Robot locomotion, obstacle courses, and driving stacks use the same idea with physics or traffic as the environment. RLHF applies it to language: a reward from preferences, then policy updates so replies score higher and harmful ones score lower. Sutton and Barto's textbook is the standard map: agent, environment, reward, policy. AlphaGo and RLHF both sit on that loop.