What game should artificial intelligence take on next?
Source: Sergey Soldatov
This week, Google’s AlphaGo beat a grandmaster at the complex game Go – an artificial intelligence milestone (see “How victory for Google’s Go AI is stoking fear in South Korea“, “Machines are teaching themselves to grapple with the real world” and “Humans strike back: How Lee Sedol won a game against AlphaGo“). Here’s what the experts say AI’s next big challenge should be.
No-limit poker: Go represents the ultimate in games where all the information is available to the players. But AI still struggles with games where information is incomplete – like poker, where a player doesn’t know what card is coming next.
“Computers have beaten the best people at heads-up limit Texas Hold’em, but not yet at no-limit, a much more complicated game,” says Peter Stone at the University of Texas at Austin.
Diplomacy: “AlphaGo doesn’t know the meaning of any of the symbols it is so adroitly manipulating: it doesn’t even know that it is playing Go,” says Mark Bishop at Goldsmiths, University of London. So he suggests the strategy board game Diplomacy, in which players pose as European powers competing for land and resources.
Diplomacy embodies many of the obstacles between current and true AI. “Interestingly, it is a game that in theory a computer could play very well, as moves are communicated in writing,” says Bishop. But it would have to pass the Turing test – humans could team up against the AI if they figured out which player it was.
“These twists on gaming go beyond the mathematical challenges being breached by current AI”
StarCraft: In Go, there might be about 300 possible moves at any time. In StarCraft, a strategy video game with hundreds of pieces, there might be 10300. “You can’t even examine all possible moves in the current state, let alone all possible future move sequences,” says Stuart Russell at the University of California, Berkeley.
Instead, the AI would have to consider its actions and goals on a higher level, then work out a plan to get there – requiring reasoning methods applicable to a wider range of real problems.
Dungeons & Dragons: “What we’re seeing with AlphaGo is not trying to prove or disprove a humanlike sense of reality or believability, but instead is purely goal-centred – to win the game,” says Julie Carpenter at the University of Washington in Seattle. She says it would be interesting to throw AI at something like a role-playing game. There, the machine’s goals wouldn’t be as obvious. It would need to rely on skills like social communication and higher-level situational awareness in order to succeed.
Cheating: Human players can read their opponent’s faces and body language for clues about what to do next. They can also get ahead by using deceptive tactics, like misdirection. Could a robotic hustler ever successfully spot these false behaviours – or even cheat without being detected? “These twists on gaming go beyond the largely mathematical challenges that are currently being breached by current AI,” says Ronald Arkin of the Georgia Institute of Technology in Atlanta.
The real world: “I’m not particularly interested in seeing AI pitted against other games,” says Murray Shanahan at Imperial College London. That’s useful for testing an algorithm or new learning methods, he says, but the true frontier is the real world. “When machine learning is as good at understanding the everyday world as it is at Go, we’ll be well on the way to human-level artificial general intelligence.”
This article appeared in print under the headline “Time to raise the game”
| }
|