Reinforcement Learning

Reinforcement learning takes a fundamentally different approach from other forms of machine learning. Instead of learning from a dataset of examples, an agent learns by interacting with an environment and receiving rewards or penalties for its actions. Think of it like training a dog: you don't show the dog a thousand examples of sitting; you reward it when it sits and withhold the reward when it doesn't. Over time, the agent learns which sequences of actions lead to the best outcomes. This approach has produced some of AI's most celebrated results - AlphaGo beating the world champion at Go, systems learning to play Atari games from raw pixels, and robots learning to manipulate objects through trial and error. The strengths are genuine: reinforcement learning can discover strategies that humans would never think of, and it excels in situations where the optimal approach is not known in advance. The limitations are equally real: reinforcement learning typically requires enormous amounts of interaction with the environment, which is fine in a video game but impractical in contexts where each trial is expensive or risky. Training is often unstable and sensitive to how you define the reward. For most business applications, supervised learning remains more practical, but reinforcement learning has a growing role in areas like robotics, recommendation systems and resource optimisation.