Reinforcement learning (RL) is a machine-learning paradigm in which an agent learns a policy-a mapping from states to actions-by interacting with an environment and optimising for cumulative reward over time. Unlike supervised learning, there are no labelled examples; the agent must discover effective strategies through trial, error, and the feedback signal provided by a reward function.

Think of it like training a dog: you do not hand the dog a manual on how to sit. Instead you give treats when it does the right thing and withhold them when it does not. Over thousands of repetitions the dog infers which behaviours earn rewards without ever being explicitly taught the rule.

A classic RL benchmark is teaching an agent to play the Atari game Breakout. The state is the pixel grid, the actions are paddle movements, and the reward is the game score. Over millions of episodes, algorithms such as DQN learn to achieve superhuman performance purely from this signal-no human-labelled training data required. In robotics, RL trains physical arms to grasp objects by rewarding successful picks and penalising drops.

RL underpins RLHF, where human preference ratings replace or supplement a hand-coded reward signal to align LLMs; it also connects to Loss Function design and Gradient Descent optimisation used during policy updates. Concepts like the Reward Model and reward hacking are direct extensions of core RL ideas.

You will encounter reinforcement learning when reading about game-playing AI, robotic control systems, and-increasingly-the alignment and fine-tuning pipelines behind modern large language models.