Google’s DeepMind, Blizzard Create A StarCraft II AI Toolkit

August 10, 2017 - Written By Daniel Fuller

StarCraft II and other deep, strategic video games can help you learn to think on your feet and sharpen your split-second decision skills and reflexes, but thanks to a collaboration between DeepMind and Blizzard, that principle can now also be applied to train AI in a variety of ways. The popular game calls for world-class skills at the top levels, to the point that it’s become part of the eSports ethos, which makes it a suitable playground for AI. Blizzard and DeepMind’s special toolset is available to the public as an open source project, meaning that anybody who has a decent computer and a copy of StarCraft II can now try their hand at building and training an AI program.

The research and development taking place in this operation centers on SC2LE, a specially made set of tools crafted around StarCraft II. The core of it all is a special machine learning API that was tailor-made by Blizzard, and can hook an AI into the game. Beneath that is an open-source implementation of DeepMind’s own AI toolset, called PySC2. Those who download SC2LE will also find an anonymized gameplay dataset, which currently holds about 65,000 game replays. That number is going to increase in the next few weeks, and the plan is to have it go above half a million. To help test the AI, there are a few simple mini-games included that work within the bounds of vanilla StarCraft II, though researchers that want to use mods or create their own mini-games can easily do so using PySC2 and Blizzard’s machine learning API. A joint paper from Blizzard and DeepMind is also in the package, detailing the creation of the toolset, and early results using it for AI training and testing. The PySC2 toolset essentially breaks that game down into layers, which the AI hooked into the machine learning API can translate to input and output for risk and reward scenarios, analyze long-term effects, and generally play around with to learn the game, thus growing its cognitive arsenal by applying and synthesizing everything it has learned.

StarCraft and its sequel are prime stomping ground for AI due to several reasons; the gameplay is a multi-faceted, drawn-out affair, and requires long-term goal-setting and decision-making in order to truly succeed, as well as good memory, since only part of the map is visible at any given time. The glut of players in both games competing online at high levels also gives researchers tons of replay data to use for training AI. The first StarCraft game is actually already being used by AI researchers, but StarCraft II is far more complex thanks to its more modern engine which makes the possibilities for actions, goals, and game outcomes practically endless.