Reinforcement learning — Glossary

Reinforcement learning (RL) trains an agent that observes states, takes actions, and receives rewards. Over time the agent improves a policy to increase expected return. Challenges include exploration, delayed credit assignment, and reward misspecification.

RL appears in games, robotics research, operations problems, and post-training methods for language models (such as optimizing against a preference-based reward model). Not all modern AI training is RL; much large-scale pretraining remains self-supervised prediction.