A Gentle Intro to Reinforcement Learning

A different paradigm

Unlike supervised learning, reinforcement learning (RL) has no labeled answers. An agent takes actions in an environment, receives rewards, and learns a policy that maximizes long-term reward.

The vocabulary

• State — the current situation.
• Action — what the agent can do.
• Reward — feedback signal.
• Policy — the strategy mapping states to actions.

Where it shows up

Game-playing (AlphaGo), robotics, and — importantly for LLMs — RLHF, where human feedback shapes a model's behavior. RL is powerful but sample-hungry and tricky to tune; start in simple simulated environments like Gymnasium.