The announcement at BlizzCon 2016 that met with the most muted response was arguably the most revolutionary.
While new content for the likes of Hearthstone, Heroes of the Storm, Overwatch, and Diablo III drew appreciative roars from the Blizzard faithful, the news that Googleâ€™s DeepMind branchâ€”which is dedicated to developing sophisticated Intelligenceâ€”would be teaming up with the makers of Starcraft 2 to further its research on AIÂ elicited more of a murmur.
Perhaps the lack of enthusiasm was down to taste. After all, why would the plans of AI scientists be of interest to Starcraft 2 players? As it turns out, if the collaboration between DeepMind and Blizzard isÂ what its developers hope it could be, players will see very tangible benefitsâ€”and so will many others outside the video game space.
AIÂ has been an integral part of video games for quite some time now, and as Starcraft 2â€™s executive producer Chris Sigaty points out, the AI that DeepMind hopes to develop wonâ€™t be the first to play the Blizzard RTS. Currently the games uses a scripted AI. It’s created by a team of designers andÂ engineers thatÂ have an understanding of the way the game plays, with theirÂ expertiseÂ placed into a script and fed a list of conditions. For example, if X amount of resources is collected, the AIÂ should spend the value of Y on Z units.
"Theyâ€™re elaborate and interesting to play against; thereâ€™s some random choice and specific choices that happen," says Sigaty. "They can expand on their terrain, gather resourcesâ€”basically do the things you need to do in Starcraft 2. But these [AI] are scripted. Theyâ€™re predictable and ultimately, unless it cheats, thereâ€™s not an AI out there who can beat a human."
This holds true for Starcraft 2â€”the best e-sports players routinely beat AI opponents. The aim DeepMind is to develop an AI that plays the game the same way as a human, potentially one that can beat the best players, which widens the potential for deep learning. And the DeepMind team hasÂ already had some success in the gaming field: in March this year, the teamâ€™s AI AlphaGo saw off the world-class Go player Lee Sedol in a best-of-five series that it won 4-1â€”leaping what was a major hurdle for AIs in the past.
It’s not (just) a game
Oriol Vinyals, one of the research scientists at DeepMind, says that Starcraft 2 may help AI research take a massive step forward, becauseÂ a game in which pieces (or in this case units) arenâ€™t static and movement isnâ€™t turn-based helps create a better benchmark for AI development.
"We feel that the complexity of Starcraft 2 enables us to now start testing algorithms that perhaps some years ago we were not ready to tackle," he says.
In Starcraft 2, much of the information a player needs to succeed in a match is initially hidden, and players are required to scout terrain while managing resources and building units in order to ascertain their opponentâ€™s position. Static AI views the movement on the board as parcels of data, and in order for DeepMind to use Starcraft 2 as a learning experience, that data has to be kept hiddenâ€”in much the same way it would be for a human player.
"The process is more complex because the game hides units and players donâ€™t know where the enemy is. It focuses on planning and resource management and then players have to factor in whether theyâ€™ve played their opponent before, and how they approach them with this knowledge," says Vinyals.
"These are all very critical AI challenges that weâ€™ll be able to tackle with this new environment. Crucially, even if it isnâ€™t immediately successful, you can start building benchmarks with how it performs, which allows us to keep pushing the envelope on the state-of-the-art which is great. Again, by making it open, everyone can add their ideas to help with the process."
The collaboration between Blizzard and DeepMind isnâ€™t a closed shop. As part of the announcement at the BlizzCon keynote, Vinyals said that Starcraft 2â€™s AI research environment would be open to any developers, hobbyists, or players who wanted to get involved. DeepMind is putting out the call in a global sense because it needs the world at large to help with this huge challenge. The API is planned for release as a patch for the game in the first quarter of 2017, and will be available to players as part of the free Starter Edition. But, as Vinyals notes, there is a lot of work to be done ahead of the launch.
"The collaboration started very recently," he says. "The first step is to build the environment, which allows us to place our [AI] agent in it and have it start to make observations while playing. Weâ€™re focused totally on building that environment because we have to release it to the public; it has to be a benchmark for other AI researchers as well. There are many people researching AI so making sure that environment is something they can all use is very important."
â€śThe next step would be introducing the agent. We haven’t tackled that yet except on a simplistic level.â€ť
http://arstechnica.com/gaming/2016/11/starcraft-2-google-deepmind-ai/ via http://arstechnica.com #CIO, #Technology