AI boffins turn to StarCraft to train future neural networks
TorchCraft to build machine-learning agents from cosmic battles
StarCraft could be the next battleground for AI, as researchers create an open framework that tests deep-learning methods in the real-time strategy game.
Teaching AI to play games is serious business. Games act like milestones; when a machine is superior to humans at playing difficult games, it’s a sign that its neural net has reached a new level of intelligence.
IBM’s supercomputer, Deep Blue, conquered chess in the late 1990s. This year, Google DeepMind achieved notoriety for creating Alpha Go, an AI that beat Go master Lee Sedol. The next game in researchers’ sights is StarCraft.
A paper [PDF], released on Thursday by eggheads at Facebook and the University of Oxford, presents a new way to test deep-learning methods on real-time strategy (RTS) games like StarCraft by creating something called TorchCraft.
The name is an amalgamation between StarCraft and Torch, a machine learning library. It helps developers build AI agents capable of handling the difficult environment of StarCraft.
The researchers connect the machine learning library to StarCraft: Brood War by injecting code in the game engine that acts as a server. The server relays information about the state of the game to the external machine-learning code and receives commands to send to the game.
There are already a number of platforms for games for the Atari 2600 plus Super Mario, Doom and Minecraft. StarCraft has been a goal for researchers, and is considered one of the most difficult games for computers to play.
Today's StarCraft bots are pretty mediocre at best, and cannot compete with professional players yet.
The goal is to take control of the map. Units are dispatched to create buildings and collect resources that can empower a player with weapon upgrades or more soldiers. At the same time, players have to prevent their areas from being invaded, as well as launch attacks on the enemy’s units.
It’s a game that requires strategic planning, decision making, and improvisation, where bots have to play over a dynamic environment that they can only partially see, making it an “excellent benchmark” for AI.
“TorchCraft is applicable to any video game and any machine learning library or framework,” the paper said.
Games have long been of interest to AI researchers, as the virtual environments are a good way to test AI that may eventually be applicable in the real world. The highly popular, violent Grand Theft Auto games are being used as testing grounds for autonomous driving software.
Scores also provide researchers with quick feedback to see if their system is performing well, and are used as a reward in deep reinforcement learning. The researchers said they will release their code soon. So watch out, human StarCraft players – AI could soon be joining a tournament near you. ®