Resistance is futile: Deepmind's AlphaStar AI hammers StarCraft II pros in world first
The Alphabet-owned AI firm, DeepMind AI, is hoping to solve some of the most complex problems that exist in the world today using a variety of methodologies. Its main tool is reinforcement learning, wherein artificial intelligence ‘agents’ compete against each other on a given task and collectively master it using cooperation and strategic thinking.
From the early days of competing at pong, in the TV game show Jeopardy and taking on world chess grandmasters, AI has graduated to the big leagues, competing in StarCraft II, arguably the most complex strategy game humanity has ever devised.
StarCraft is a real-time strategy (RTS) game which involves three distinct races (Terran, Protoss and Zerg) mining resources, building up armies and employing devious and complex strategies, often involving thousands of steps, to overcome the opponent. The game boasts more than 2.4 million players worldwide, with millions more fans watching around the world.Also on rt.com Are eSports the future of Olympics? 5 things you wouldn’t expect from pro gaming
In StarCraft, roughly 300 basic in-game interactions can yield millions and millions of possible permutations which must be calculated and adapted on an ongoing basis. Over the course of 18 months of study and testing, DeepMind’s AlphaStar AI studied roughly half a million anonymized games between human players that were provided by the game’s developer, Blizzard.
By simply copying human player strategies, AlphaStar was soon able to overcome ‘Elite’ level AI players in the game. In an internal competitive league, the AI initially favored high-risk, all-in strategies before developing a more nuanced tactical approach. The AlphaStar League ran for 14 days in total, with individual AI agents experiencing up to 200 years of real-time StarCraft play across thousands of instances of the game.
When competing, AlphaStar could see the entire field of play, unlike a human player, but the DeepMind team created as level a playing field as is currently possible. This essentially limited some of the more devious strategies the professional StarCraft players could employ.
However, the AI cannot react as quickly as a human player can and cannot execute more Actions Per Minute (APM) than a human player can; for context, most professional StarCraft players have an APM of several hundred.
Then came the AI’s first real opponents; professional, human StarCraft players Dario ‘TLO’ Wunsch (Zerg), and Grzegorz ‘MaNa’ Komincz (Protoss), of Team Liquid, who has been playing the game since he was five, were both drafted in to test their mettle against the AI. It did not go well.
“I assumed after the first match I would have a very good idea how to play against it. I did not,” TLO said after losing every match in the five-game series against AlphaStar. The same was true for MaNa.
In game one versus TLO for example, the AI managed 277 APM compared with the human player’s 559.
“Hopefully I get to have a rematch where I get to play my strongest race,” a gracious TLO said in defeat, having been forced to play against the AI as Protoss instead of his preferred Zerg.
However, there is one glimmer of hope for humanity: when the DeepMind team subjected AlphaStar to the same restrictions faced by human players of the game (the inability to see large sections of the map), instead allowing the AI to only operate using the in-game camera, MaNa was able to overcome early adversity and beat the machine in a live game streamed live on Twitch.
Real world application
While this may also sound trivial to the casual observer, this latest breakthrough in mastering incredibly complex systems using incomplete and rapidly-changing information should, in time, improve AI capabilities with regard to weather forecasting, climate modelling, and language interpretation.
3/3 While StarCraft is ‘just’ a (very complex!) game, I’m excited that the techniques behind #AlphaStar could be useful in other problems such as weather prediction & climate modeling, which also involve predictions over very long sequences. Peer-reviewed paper is underway.— Demis Hassabis (@demishassabis) January 24, 2019
“We’re very excited about the potential to make significant advances in these domains using learnings and developments from the AlphaStar project,” the DeepMind team said. An incredibly detailed explanation of the entire process, along with the replays of the games, can be viewed here.
Think your friends would be interested? Share this story!