Reinforcement Learning (RL)

← Flow Matching & Modern Diffusion Q-Learning & Deep Q-Networks (DQN) →

Learning through trial and error, an agent takes actions in an environment, receives rewards/penalties, and learns a policy to maximize cumulative reward. Behind AlphaGo and game-playing AI. The multi-armed bandit is RL stripped to a single state, where the explore-exploit tradeoff appears in its purest form.

Key terms

AgentEnvironmentRewardPolicyEpisodeMulti-armed bandits

Learn these first

← Machine Learning (ML)

Deep dive · 6 min

Reinforcement Learning

Reinforcement learning has no answer key at all. An agent takes actions in an environment, occasionally receives a reward, and must work out which of its many past actions deserve the credit or blame. That credit-assignment problem is the field's defining difficulty: the win comes forty moves after the brilliant move that caused it.

Where it earned its fame

Games gave RL clean environments and cheap experience: systems trained by self-play reached superhuman strength at Go, chess, and complex video games, discovering strategies no human had taught them. Robotics uses RL where the physics is too messy to script. The stochastic-approximation mathematics underneath traces back to Robbins and Monro in 1951, decades before anyone called it reinforcement learning.

Why your chatbot is polite

RL's most consequential deployment is inside LLMs. RLHF (reinforcement learning from human feedback) turns human preference rankings into a reward signal that shifts a raw text predictor toward being helpful and harmless. The newest wave, sometimes called reasoning RL, rewards models only when their multi-step reasoning verifiably works, on math that checks out and code that passes tests. When a model pauses to think before answering, you are watching behavior that reinforcement learning selected for.

In one glance