Reinforcement Learning
also: RL
Learning by trial and error: an agent takes actions, gets rewards, and learns a policy that maximizes long-term reward.
Reinforcement learning (RL) trains an agent to make sequential decisions by interacting with an environment: it takes actions, receives rewards, and learns a policy that maximizes cumulative reward over time. Unlike supervised learning, there are no labeled correct actions — only the reward signal.
Worked example: a game agent tries moves, gets +1 for winning and −1 for losing, and over many episodes learns which actions lead to wins — the same paradigm (RLHF) that aligns language models to human preferences. Gotcha: RL is powerful but notoriously sample-inefficient and unstable — sparse or delayed rewards make credit assignment hard, and a misspecified reward gets “hacked” (the agent maximizes the metric, not your intent) — which is why reward design and methods like PPO and GRPO matter so much.