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.
âWeâre so good at poker that not even a supercomputer can beat us,â says professional player Jason Les. âIf we lose, we will also lose that prestige.â
That could change this month. Les is one of four top Heads-Up No-Limit Texas Hold âEm players representing Team Human in the 2017 at Rivers Casino in Pittsburgh. They are taking on a computer program called Libratus, developed by researchers at Carnegie Mellon University.
Advertisement
An AI that can successfully beat poker professionals wouldnât just claim another gaming victory over humans. It would also signal an ability to work with incomplete information â a skill that has a wide range of applications beyond just poker.
The competition started on 11 January and runs for 20 days, with more than 120,000 hands being played to remove any element of chance from the results.
AI underdog
Ahead of the event, the odds were on the reigning human champions, who won the last competition against the AI in 2015. âInternational betting sites are considering the AI to be a 4:1 or 5:1 underdog,â says , one of the creators of Libratus.
At the end of day two, however, Libratus was up by $150,126. It was winning against three players and losing against one.
The AI has spent the equivalent of 15 million hours of computation honing its strategies, heading towards what game theorists call a Nash equilibrium: the point at which no further improvement is possible. Sandholm and his colleagues wonât reveal exactly how Libratus works, but they will say it hasnât been fed any particular strategy. Instead, the AI must learn the best approach for itself.
Poker is a difficult game for AI to conquer, and learning how to do so could help it develop other skills too. At any moment in a game, it can only see the cards it holds and those on the table, but not its opponentâs.
An AI could simply calculate the odds that its hand is likely to win and play accordingly â similar to how AlphaGo, the AI created by Googleâs DeepMind, beats top Go players â but knowing how to bet is less straightforward. If the AI bet high every time it had a good hand, a human opponent could easily figure this out and exploit it. The key for the AI is therefore not just working out how to play, but also figuring out how an opponent is playing.
âPoker is an excellent game for testing artificial intelligence,â says , head of the Computer Poker Research Group at the University of Alberta, Canada. âUnlike chess or Go, you never have a perfect view of the state of play, which is much more like reality.â
Another artificial challenger
Bowling is part of a team that that has beaten professional poker players. DeepStack uses a combination of machine learning and some strategic simplifications to make it possible to assess the best move in real time.
The researchers say DeepStack came out on top after playing 44,000 hands of poker against 33 professionals. Their work has yet to be peer reviewed, but it would be an impressive feat if it stands up to scrutiny.
Bowling and colleagues previously created a program to âsolveâ Heads-Up Limit poker, but Heads-Up No-Limit is a tougher challenge, as there is no restriction on the size of the bets. The number of different possibilities in a game of Heads-Up Limit is 1014, but in No-Limit itâs more like 10160.
The heads-up variant of the game, which pits players one-on-one, may not be the most popular kind of poker, but it is best suited to AI competitions. âIf there were more players, the humans would simply gang up against the AI, shifting the odds by a huge amount,â says Bowling. âNo matter how sophisticated your AI was, it would still lose.â
An AI that can excel at No-Limit poker could have applications in other tasks that require a strategy to be drawn up from incomplete information. This is the case in most real-world scenarios, such as when doctors must decide on peopleâs treatment without knowing everything about them.
âItâs inevitable that AI will eventually win,â says Les. âThat wonât be the end of poker, but it will be a massive milestone for AI.â
