<
>

Google's DeepMind AI takes on StarCraft II

play
A StarCraft-playing AI could change esports... and the world (2:50)

The DeepMind AlphaGo AI defeated Go world champion Lee Sedol this past March. The organization has now announced its next project: StarCraft II. (2:50)

It's match point at the grand finals of a huge StarCraft II tournament. One player booth contains a professional Zerg player complete with massive headphones and stoic concentration. The other booth is ... empty. After Zerg wins in the late game with an unexpected tech switch to infestor/broodlord, the crowd goes wild. The pro player comes out on stage and accepts a huge trophy. They are proud not only of winning, but of dealing a blow to their opponent the likes of which could contribute to technology benefiting millions of people worldwide.

At BlizzCon earlier this month in Anaheim, California, Blizzard announced an ambitious new project in collaboration with DeepMind, a leading artificial intelligence research company acquired by Google in 2014. After creating the AlphaGo AI that bested the world's top Go player earlier this year, DeepMind's next groundbreaking challenge will be StarCraft II. If DeepMind is able to build an AI that could learn how to beat top players such as Byun "ByuN" Hyun Woo in the complex real-time strategy, tactics and resource management of this game, it would be a giant step forward in AI research. And with DeepMind's interest in using its research to solve hard problems in areas such as healthcare and energy efficiency on a massive scale, this Starcraft II project could impact the whole world.

Soon after AlphaGo's Go victory, there were signs that DeepMind would take on StarCraft next. This was not lost on legendary StarCraft player/commentator and former competitive chess player Dan "Artosis" Stemkosi, for whom StarCraft seemed like the logical next step for AI research after games like chess and Go. While he thought that an AI could have a huge advantage in terms of mechanical skill and precise resource calculations, he also thought this advantage would not be enough to defeat experienced, resourceful pros.

"I think this is a good place to bring up the super-confident statement made by The Emperor of Terran himself, Lim "BoxeR" Yo Hwan: 'Even if it has studied all of the many strategies I've used, I'll go at it with an unstoppable strategy I've prepared,'" Artosis wrote.

"As soon as I saw this quote, I knew exactly what he meant. There will be some flaw within the AI. Everything you do in StarCraft has a cost, somewhere you have to cut a corner. With enough games under his belt, BoxeR, or any other extremely experienced pro gamer, would find where the AI would be 'cutting' these corners. ... There will be patterns, and patterns are dangerous in StarCraft."

StarCraft play is highly situational, the argument goes, and trying to manage all of these pieces in real time based on certain algorithms or formulas cannot top humans in the realm of intuitive pattern recognition (and breaking such patterns). In the long-running AIIDE StarCraft: Brood War AI Competition, the three best programs of 2015 each faced off with a top Russian StarCraft player. They all lost.

Furthermore, DeepMind's proposed AI is at an interesting disadvantage compared with, say, IBM's Deep Blue. Deep Blue had sophisticated pre-programmed rules and calculations that allowed it to dominate in chess, but it couldn't do anything else; this is called a "narrow AI." DeepMind's StarCraft AI will be the opposite. Its goal is to become a "general-purpose learning machine," which means it will learn the game from scratch, with only very basic game-specific information guiding it through automated trial-and-error in order to develop real, generally applicable intelligence. Essentially, this digital mind will cut its baby teeth on conquering StarCraft in order to learn how to adapt to and operate in the outside world.

But just like Legend of Zelda's Link needs the Old Men to provide him with hints and weapons to achieve success in Hyrule, a true StarCraft AI also needs information and tools to bootstrap itself from blank slate to World Championship Series winner. The DeepMind team has Google's massive computational infrastructure to help with one side of it, but that's not enough. It needs the community.

This is where Blizzard comes in. In a BlizzCon session fittingly hosted by Artosis, StarCraft II senior software engineers described two ways in which the game itself would enable players and fans everywhere to participate in this grand experiment. First, all players from bronze leaguers to pros will be feeding massive amounts of data into DeepMind and other developing AIs just by generating replays on ladder. Second, Blizzard is creating a powerful set of software tools that will enable modders and researchers to develop AIs that can fully interact with the game. Blizzard is transforming StarCraft II from a popular computer game and esport into a platform for cutting-edge computer science all over the world.

This unprecedented partnership is coming at a perfect time for game development and AI research.

"On a large, competitive scale, there has not really been any other scenario ... where the game companies have actively helped [third parties to create game AIs], and you really need that level of assistance," Matthew Fisher, an AI researcher who recently finished his Ph.D. from Stanford, said in an interview with ESPN.

In 2011, Fisher was building a StarCraft II AI with a similar approach to DeepMind, and he had to do it all on his own.

"The main reason you're seeing this particular thing happening now is [because] it's the first time the artificial intelligence community has really been developed enough ... to the point where you [could start to see AIs that] would challenge professional players," he said.

Sure, blowing the AI research field wide open is a big deal, but how could this project affect the esports scene? Polish StarCraft II pro player Artur "Nerchio" Bloch, a Red Bull-sponsored player on EURONICS Gaming who won DreamHack Valencia earlier this year, spoke to ESPN about his perspective.

"Some people were saying it could be used to make some kind of dynamic tutorial for players ... but I don't think that's the way to go, really," he said. "If anything, this AI is going to help us to develop a lot of strategies ... we haven't seen before."

Nerchio studied information technology in school and happens to have a particular interest in artificial intelligence. He's very excited for the DeepMind project. For all that, the larger impact of a theoretical AI that could defeat top players like him is unpredictable.

"I'm a little bit worried, I guess, because an AI beating the best players in the world is always exciting, but at the same time it can be damaging to the competition or to the community surrounding the games, knowing that there's something better than the best players out there," Nerchio said. "I am excited but worried at the same time."

Although the specter of an unbeatable AI might loom large, Fisher believes the process of getting to that theoretical point could produce some of the most interesting and exciting StarCraft ever.

"Expert players will find weaknesses in the AIs, and then you'll have this really cool period of a year or more where the AIs are adapting, Google is updating its AI to deal with the power of expert players, expert players are battling against Google's AI ... [in] this very cool interplay," Fisher said.

The benefit of the research to StarCraft is not ultimately focused on esports, however.

"I feel that, for the players of the game, this should be equally exciting regardless of your level," DeepMind research scientist Oriol Vinyals said to ESPN. "It may be more obvious that the esport scene may be more enthusiastic about it, but in fact I believe that 'average players' may see some benefits sooner, such as AI coaches, or additional aids for the novice players so they can focus on certain aspects of the game (e.g., by letting AI control your units and letting the player decide on the high level strategy)."

We don't know what will happen with the Google DeepMind StarCraft II project -- whether their AI will be able to defeat pro players, whether it will be able to revolutionize player coaching or whether it will be able to transfer its approach to the real world at all. But people won't give up, because the desire to change the world through technology and the fascination with a game that ruthlessly challenges us to think are inextricably intertwined in a new generation.

"As a teenager, I played the original Starcraft and Brood War for a while. I was a very good player on the Spanish and European scene (top 32 or so, I'd say) back in around 2000," Vinyals said. "Later on, as an AI Ph.D. student at Berkeley when I was approached to be the 'coach' of the bot we built back then, I had no doubt that my unique skills in the game plus knowledge about AI would be a nice mix to build what we did. It is extremely nice that, even further into my career, StarCraft is making yet another comeback."