3.6 Introduction to Reinforcement Learning

1. Introduction to Reinforcement Learning

In this video David explains the principles behind reinforcement learning, an approach where a machine learner is trained using feedback from the environment rather than a dataset.

Introduction: what is the problem which reinforcement learning is trying to solve?



Reinforcement Learning: rewarding positive outcomes

Linking to Images - Convolutional Neural Networks, analysing complex environments



Summary: what is reinforcement learning and when should we consider applying it?



In the video we are actually looking at the problem of object classification (given an image, can we tell what it is). This is actually not sufficient for more complex scenes (like the game of DOOM), where we want to not only classify multiple objects, but also understand where they are in the scene. There are a range of techniques for doing this, including Region CNNs (R-CNNs) and YOLO methods.

Finally you may have heard of Google Deepmind’s AlphaGo, one of the most famous AI systems of recent years. AlphaGo was the first AI system to be able to defeat professional Go players (Go is a very open game, and was considered unassailable using traditional AI techniques). AlphaGo first trained on supervised learning to replicate human Go players, then switched to reinforcement learning to hone its game and improve. A later version, called AlphaGo Zero, used no training data at all - and was entirely self taught. 

The work on AlphaGo really shows the need to use reinforcement learning if the objective is to surpass human performance.