Reinforcement Learning (RL)

An area of machine learning concerned with how software agents ought to take actions in an environment in order to maximize the notion of cumulative reward.

How it works

In RL, an agent observes the state of an environment, selects an action, and receives a reward signal. The agent's goal is to learn a policy — a mapping from states to actions — that maximises cumulative future reward. Algorithms like PPO and GRPO use sampled trajectories to estimate how much better or worse each action was compared to the baseline, then update the policy to make better actions more likely. The agent learns entirely from trial and error, requiring no labelled examples.

Why it matters

RL is the training paradigm that produced some of AI's most dramatic results: AlphaGo defeating world champions at Go, OpenAI Five mastering Dota 2, and robots learning to walk and manipulate objects. In LLM training, RL (via RLHF) is what transforms a raw language model into a helpful assistant by rewarding responses that humans prefer. DeepSeek-R1 showed that RL alone can elicit sophisticated reasoning from a base model, without supervised imitation.