Imitation and Transfer Learning

At the time of writing, a new AI called AlphaStar, a deep reinforcement learning (DRL) agent, used imitation learning (IL) to beat a human opponent five-nil playing the real-time strategy game StarCraft II. AlphaStar was the continuation of David Silver and Google DeepMind's work to build a smarter and more intelligent AI. The specific techniques AlphaStar used to win could fill a book, and IL and the use of learning to copy human play is now of keen interest. Fortunately, Unity has already implemented IL in the form of offline and online training scenarios. While we won't make it to the level of AlphaStar in this chapter, we still will learn about the underlying technologies of IL and other forms of transfer learning.

In this chapter, we will look at the implementation of IL in ML-Agents and then look to other applications of transfer learning. We will cover the following topics in this chapter:

IL or behavioral cloning
Online training
Offline training
Transfer Learning
Imitation Transfer Learning

While AlphaStar performed a stunning tactical victory against a human pro player in an RTS game, it has still come under scrutiny for the type of play and actions it used. Many human players stated that the AI's tactical abilities were clearly superior, but the overall strategy and planning were abysmal. It should be interesting to see how Google DeepMind approaches this criticism.

This will be an exciting chapter, and will provide you with plenty of training possibilities for your future developments, which all starts in the next section.

Table of Contents for Imitation and Transfer Learning

Create new playlist

Sign In

Sign Up

Table of Contents for
Imitation and Transfer Learning