Introduction to reinforcement learning
In previous chapters, we discussed a model that learns from a large amount of text. Humans—and increasingly, AI agents—learn best through trial and error. Imagine a child learning to stack blocks or riding a bike. There’s no explicit teacher guiding each move; instead, the child learns by acting, observing the results, and adjusting. This interaction with the environment—where actions lead to outcomes and those outcomes shape future behavior—is central to how we learn. Unlike passive learning from books or text, this kind of learning is goal-directed and grounded in experience. To enable machines to learn in a similar way, we need a new approach. This learning paradigm is called RL.
More formally, an infant learns from their interaction with the environment, from the consequential relationship of an action and its effect. A child’s learning is not simply exploratory but aimed at a specific goal; they...
