Poker? Done that. Now the next challenge…
Poker, as has previously happened to chess and Go, has joined the games that a set of algorithms is already capable of playing better than the human champions can manage. On January 31, after twenty days of Heads Up, No Limit Texas Hold ’em, four people considered among the best professional poker players in the world were defeated by an artificial intelligence machine, Libratus, the product of the work of researchers of Carnegie Mellon directed by Tuomas Sandholm.
Twenty days watching computer screens, playing about 120,000 hands, and meeting at night in their hotel rooms to coordinate joint strategies were not enough to beat an algorithm that quickly understood the strategies employed by humans and it soon overcame them.
The game was clearly dominated by Libratus from the first moment: the human players were not even close to winning at any time. The aim of keeping the championship going to the end was to achieve a victory that could be considered statistically significant, that is to say, winning 99.7% of the time is hardly the product of chance.
What really matters here is that the algorithms used were not specific to the game of poker, nor did they try to exploit the mistakes of the Libratus’s opponents. They simply took the rules of the game as their inputs and focused on improving their own strategy by taking into account the cards dealt, those on the table and the bets placed by each player. Texas Hold ’em, with its unlimited betting and the uncertainty of two hidden cards on whose potential values player speculate, offers a very good example of imperfect information play, and serves as an appetizer for other non-gambling activities such as negotiation, cybersecurity, finance, or even research on antiviral treatments (taking the mutations of the virus, whose genetic sequence is known, as uncertain variables that allow it to survive certain drugs).
There are plenty of areas similar to poker: we’re no longer speaking about a machine that can learn the rules of a game and apply computational brute force to calculate. What Libratus’s victory means in simple terms is that artificial intelligence is better at making strategic decisions based on uncertain information than humans are.
If you thought that a machine was only capable of repeating what it had been programmed to do, think again: a machine has been able to analyze 120,000 poker moves and, given the cards dealt it, the cards already on the table and the bets of each of its opponents, consistently won on a statistically significant number of occasions, enough to rule out luck or chance.
So next time you sit down to play a hand of poker, remember that no matter how well you do, there is a machine out there that will always beat you.
And from now on, that won’t just apply to card games…
(En español, aquí)
