When it comes to StarCraft, artificial intelligence has some catching up to do when taking on humans.
This became clear recently when four separate bots were defeated by Song Byung-gu, the professional StarCraft player, in the very first contest between pros and AI systems in live matches of the popular video game. One bot, "CherryPi", was born in Facebook's very own AI research lab. The other three bots represented Korea, Norway, and Australia.
Man vs machine
The competition was staged at Sejong University in Seoul, Korea. The venue has played host to StarCraft AI contests since 2010. As opposed to AI systems taking on humans, those previous contests pitted AI systems against one other and were partly organised by the Institute of Electrical and Electronics Engineers (IEEE), an engineering association based in the U.S.
Whether humans vs AI or AI vs AI, such contests have managed to capture the attention of a curious public. Of course, humans vs humans remains the core of competition and has some way to be beaten in terms of pure entertainment value, but many industries have pitted virtual bodies against one another, not just for entertainment, but in gambling. Virtual horse racing, tennis and football has been a popular strand of gambling online for years, and it is very much treated the same as real sport, with odds set up in a similar way and bookie affiliate sites offering benefits such as free bets and deposits for use on either human sports events or virtual sport. Whether watching virtual reality gaming or sport becomes as popular as gambling on it remains to be seen.
From AlphaGo to StarCraft
While it failed to come under the same level of worldwide scrutiny as last year's contest between a Go champion and Alphabet's AlphaGo, the recent contest is not without significance, as StarCraft is considered a game that is difficult for AI to get to grips with. After Lee Sedol fell easily to AlphaGo, along with a number of other impressive displays by AI in Atari games and chess, the AI community shifted its attention to whether AI was also capable of defeating their human counterparts in such real-time games as StarCraft.
Go enabled humans and bots to view the main board and invest time in devising a winning strategy. StarCraft, on the other hand, asked players to plan ahead, formulate a strategy, and test their recall ability, all at the same time, inside a simulated and constrained world. This is why StarCraft is considered by researchers to be a more efficient game to help advance AI.
Sejong's 27 minutes of fame
If it were staged now, the contest wouldn't be much of an event. During the contest in Sejong, Song destroyed all four of the bots in under 27 minutes. That was the case even though all four bots had the ability to multi-task and generally move at greater speed. There was one point in a game when the Norwegian bot was completing an impressive 19,000 actions every minute. Most pros aren't capable of making hundreds of moves in that time.
Not enough data for bots
The contest was organised by university's professor in computer engineering, Kim Kyung-Joong, who said that AI was partly hindered by a lack of widely accessible StarCraft training data. He pointed out that AlphaGo was able to improve its ability to be competitive by learning data related to the game. That is set to change, however, as in August, AI company DeepMind, along with game developers Blizzard Entertainment, released a number of AI development tools compatible with StarCraft 11. It would be interesting to see how bots would use that data in the likely event of a rematch.
