Facebook’s AI Team Trying to Verify Google’s Results

Despite Google releasing their paper on DeepMind, which they used to create a superb bot capable of beating any human Go player, they didn’t spill all their beans. Now, Facebook’s AI team is releasing an open source initiative to redevelop those results.

In the 90’s, IBM was able to create DeepBlue, a chess AI that used next-move-prediction to beat the world’s best players. By creating a tree of all possible moves, DeepBlue could think many steps ahead of what a human could and used this to embarrass the human grandmasters of the time.

While Chess was beaten with AI, the game of Go would take another two decades to solve because Go has a lot more pieces and moves available, making the simplistic move-prediction technique unfeasible to implement.

Google didn’t solve this by just waiting for hardware to catch up — move-prediction for Go still isn’t feasible with modern hardware just as it wasn’t decades earlier. Rather, they worked on an entirely different approach with their algorithm which they named DeepMind.

An implementation of this algorithm called AlphaGo describes a process for allowing a computer to self-generate optimal moves.

AlphaGo Overview

AlphaGo Zero Cheat Sheet / Applied Data Science

Previously, we linked to an author who attempts to break down how AlphaGo Works. The essence is that there is code to allow a computer to self-generate an optimal play style for the game of Go. Rather than hard coding rules like AI of past yonder, the code here describes the rules of the game and then asks itself how it can get the most amount of points using these rules. After allowing the code to bake on powerful GPUs for thousands of hours, an optimal play strategy capable of beating all humans emerges.

Despite releasing a dense paper on how it works, Google didn’t include their actual data model that was developed with all their hours of compute time.

To make up for this lack accessibility, Facebook’s AI team has decided to open up their own implementation of a ML/AI solution to the game of Go.

Using their own ML/AI framework made specifically for games, where they are currently working on solutions for Starcraft in addition to Go, Facebook has open sourced their code as well as training data, including instructions on how to generate your own training data with nothing but a lot of GPUs and time.