NEWS & TECHNOLOGY
AI takes on top poker players Mastering the game could help AI with real-world tasks, finds Timothy Revell
TIM KAULEN/CARNEGIE MELLON UNIVERSITY
A NEW game has been added to the grudge match between humans and artificial intelligence. AI has already mastered chess and Go, but in poker everything is still to play for. “We’re so good at poker that not even a supercomputer can beat us,” says professional player Jason Les. “If we lose, we will also lose that prestige.” That could change this month. Les is one of four top players representing humans in the 2017 Brains vs AI poker tournament at Rivers Casino in Pittsburgh. They are each taking on a program called Libratus at Heads-Up NoLimit Texas Hold ‘Em, which comprises one-on-one games “Unlike chess or Go, you with no limit on the size of bets. never have a perfect view The 20-day contest involves of the state of play, which 120,000 hands. is much more like reality” Although bots are common on online poker sites, an AI that can truly master the game is a tough information – a skill with a wide challenge. Ahead of the event, the range of uses outside poker. odds were on the reigning human Poker is a difficult game for AI champions, who won the last to conquer as players can only see competition against the AI in the cards they hold and those on 2015. “International betting sites the table, but not their opponents’ are considering the AI to be a 4:1 (see “Artificial intuition”, below). or 5:1 underdog,” says Tuomas “Unlike chess or Go, you never Sandholm, one of the Carnegie have a perfect view of the state of Mellon University researchers play, which is much more like who created Libratus. As New Scientist went to press, however, ARTIFICIAL INTUITION the AI was up by over $50,000 AI has already conquered games of after almost 28,000 hands. great complexity, from IBM’s Deep The AI has honed its strategies Blue supercomputer beating Garry over the equivalent of 15 million Kasparov in a chess match in 1997 hours of computation. Sandholm to DeepMind’s AlphaGo winning a and his colleagues won’t reveal tournament against top Go player exactly how it works, but they Lee Sedol last year. So what makes say it hasn’t been fed a particular poker such a challenge? strategy and must, instead, While Go is incredibly complex, learn the best approach for itself. both players can see all the pieces An AI that can beat poker in play and use this to inform their professionals wouldn’t just claim moves. This is called a game of another gaming victory over perfect information. In poker, humans. It would also signal an ability to work with imperfect 8 | NewScientist | 21 January 2017
–Battling for humans–
reality,” says Michael Bowling, head of the Computer Poker Research Group at the University of Alberta, Canada. Bowling is part of a team that created an AI called DeepStack that they say can consistently beat professionals at the one-on-one Heads-Up No-Limit form of the game – although the work is yet to be peer-reviewed. DeepStack uses machine learning and some strategic simplifications to assess the best real-time move to make. The researchers say DeepStack
however, players’ cards are hidden from each other, so it is a game of imperfect information. An AI could simply calculate the odds that its hand will win and play accordingly, but knowing how to bet is trickier. If the AI bet high every time it had a good hand, a human player would figure this out and exploit it. The AI therefore has to not only work out how to play, but also figure out its opponent – and balance computation with something akin to “intuition”.
came out on top after playing 44,000 hands against 33 pros. Previously, Bowling and his colleagues created a program to “solve” Heads-Up Limit poker, but Heads-Up No-Limit is a tougher test as betting is unrestricted. The number of possibilities in a game of Heads-Up Limit is around 1014, but in No-Limit it’s more like 10160. The heads-up variants are best suited to AI competitions. “If there were more players, the humans would simply gang up against the AI, shifting the odds by a huge amount,” says Bowling. “No matter how sophisticated your AI was, it would still lose.” An AI that can excel at No-Limit poker could have applications in other tasks that require a strategy to be drawn up from limited information. This is the case in most real-world scenarios, such as when doctors must decide on people’s treatment without knowing everything about them. “It’s inevitable that AI will eventually win,” says Les. “That won’t be the end of poker, but it will be a massive milestone for AI.” n