How rewards teach reinforcement learning agents to behave

In reinforcement learning (RL), a software agent learns through trial and error. When it takes the desired action, the model receives a reward.

Over time, the agent works out how to execute the task to optimize its reward.

The technique can be applied to a vast array of tasks, from controlling autonomous vehicles to improving energy efficiency. But its most celebrated achievements have come in the world of games.

In March 2016, the Reinforcement Learning technique had a landmark moment. 

A DeepMind system called AlphaGo became the first computer program to defeat a world champion in Go, a famously complex board game.

The victory was reportedly watched by over 200 million people.

AlphaGo learns the game from scratch by playing against different versions of itself thousands of times, incrementally learning through a process of trial and error, known as reinforcement learning. This means it is free to learn the game for itself, unconstrained by orthodox thinking.