Chapter 09 · Sustainability & AISustainability & AI

Reinforcement learning

Meaning statusEstablishedSource recordDirect document linkedWhy these are different

Definition

A machine-learning paradigm in which an agent learns what actions to take by trial and error, receiving rewards or penalties rather than labelled examples.

References

MIT PressSutton & Barto — Reinforcement Learning: An Introduction, 2nd ed. (2018)

This source provides part of the technical or institutional basis for the definition.

Google DeepMindSafety-first AI for autonomous data-centre cooling (2018)

This source supports the explanation of how the term is applied, measured or governed in practice.

Overview

What it means

Instead of learning from a fixed dataset, the agent interacts with an environment and optimises long-term reward. Sutton and Barto's textbook defines the field. Because training requires many interaction cycles, agents are usually trained in simulation before deployment.

How it is used

Energy is the flagship sustainability application: reinforcement-learning controllers have been used for data-centre cooling optimisation — Google reported cutting cooling energy substantially after handing control to such a system — and for grid management, battery dispatch and building control.

Why it matters

Wherever a physical system must be continuously steered toward efficiency — a grid, a chiller plant, a battery — reinforcement learning is a candidate tool. The same autonomy raises verification and safety requirements before real-world control is granted.