Reinforcement Learning: Definition and Examples

Reinforcement Learning (RL) is a machine learning paradigm where a software agent learns to act in a given environment by maximizing a cumulative notion of reward. Unlike supervised learning where labeled examples are provided, the RL agent discovers the best strategies on its own through trial and error.

The functioning is based on a fundamental cycle: the agent observes the state of its environment, chooses an action, receives a reward (positive or negative), then observes the new resulting state. Over thousands or millions of iterations, the agent develops a policy — a strategy that maps each state to the most advantageous action. Algorithms like Q-Learning, SARSA, or PPO (Proximal Policy Optimization) make it possible to optimize this policy.

RL has experienced spectacular growth thanks to landmark achievements: DeepMind's AlphaGo that beat the world champion of Go, or language models like ChatGPT that use RLHF (Reinforcement Learning from Human Feedback) to align their responses with human preferences. This technique is also at the heart of robotics, autonomous vehicles, and optimization of complex systems.

In prompt engineering, understanding RL is essential because it explains why current language models behave as they do. RLHF is the reason why a LLM prefers to give helpful, honest, and harmless responses rather than simply completing text. This understanding allows one to better formulate prompts by taking into account the biases and behaviors induced by reinforcement training.