'AI still has long way to go'
Professional StarCraft gamer Song Byung-gu plays against an artificial intelligence (AI) program, in a match at Sejong University in Seoul, Tuesday. / Korea Times photo by Lee Min-hyung
Decision-making key to success for AI player
By Lee Min-hyung
No one would have ever expected the strategy game "StarCraft" to become the talk of research at the world’s most renowned tech companies such as Google, Facebook and Alibaba.
Google picked the real-time online game as its next target after successfully drawing massive attention last year by arranging go matches between its artificial intelligence (AI) system AlphaGo and human champion Lee Se-dol.
The key to victory in the game lies in how to constantly reshape strategies in various areas such as unit management, expansion and combat timing. Because of the real-time nature of the game, most critics predicted AI players would not be able to defeat human beings for the time being.
A few months after the AlphaGO matches, DeepMind, a Google subsidiary and the AlphaGo developer, unveiled its plan to hold matches in "StarCraft II," with no confirmed timelines. StarCraft developer Blizzard also welcomes the upcoming AI versus human competition by making public the game’s application programming interface (API), which makes it easier for more developers and tech players to train their own AI bots.
Korea is also no exception on the issue, as the original title is one of the most successful and loved online games here.
In a bid to jump on the AI bandwagon, Sejong University in Seoul held the world’s first official StarCraft match between its MJ bot and former human world champion, Song Byung-gu, Tuesday.
Seen above is a recent StarCraft match between former world champion Song Byung-gu and Facebook's "CherryPi" AI bot. / Korea Times photo by Lee Min-hyung
No AlphaGo shock again
As expected, the much-hyped competition ended up in a landslide victory of Song who clinched a top spot at the World Cyber Games 2007.
Song, 29, also competed against AI world champion, ZZZKBot from Australia, Norway’s TSCMOO and CherryPi developed by Facebook. After defeating MJ bot with ease, the top gamer continued his winning streak against all the three remaining AI players across the globe. It took less than 30 minutes for him to defeat all four AIs.
“I was somewhat nervous before the first match against MJ bot, so it took relatively longer time for me to compete against the Sejong University-developed program,” Song said in a post-match interview, Tuesday. He defeated MJ bot in about 10 minutes, with the remaining three competitions ending in less than 6 minutes.
“But I came to realize the AI system has loopholes in terms of decision-making in combat. Playing styles for all the AI bots were structured and stereotyped. But the level of MJ bot was beyond expectations, as it prepared for my attacks and formed its troop in a systemized way.”
However, he acknowledged that all the victories were in part due to the fact the StarCraft AI industry is at the entry level.
“Only a few companies and research facilities are investing in developing more sophisticated and human-like StarCraft-playing programs,” he said. “If Google, Facebook and other emerging tech players put in more manpower into their StarCraft AI research, I believe humans may lose matches against computer programs in less than three to four years.”
After the matches, Kim Kyung-joong, a computer engineering professor at Sejong University, said the result was anticipated because of the MJ bot’s lack of deep learning capability. Kim is heading the MJ bot development team.
“The MJ bot was developed on a script-based input system, and this is one huge difference from self-learning AlphaGo. MJ bot can operate pre-ordered tasks but cannot operate any real-time and decision-making processes.”
“Our team tested the AI bot against 20 human players before the matches, and MJ bot lost most of the practice games,” he said. “Even if the AI versus AI StarCraft competition takes places each year, our team wanted to push for a new challenge by holding matches between AIs and people.”