Ah, man vs. machine. No one can forget the worldwide attention that duels between chess grand-masters and software garnered in the late 1990s and the early 2000s. The craze included matches such as Garry Kasparov vs. Deep Blue, Vladimir Kramnik vs. Deep Fritz, and the Man vs. Machine World Team Championships in Bilbao.
In the mid-2000s, keeping in line with this trend, researchers began developing artificial intelligence for poker. As the poker boom was at its peak during this period, with the number of people playing online growing year-on-year, this came as no surprise. However, since the deep learning techniques applicable to chess did not translate to the imperfect information setting of poker, the process was a bumpier road than anticipated.
Over the years, with each set of incremental algorithmic advancements, we have reached a point where humans can no longer compete with computers. Thus, poker AI is the uncrowned champion of the game.
Top Professional Poker Players Are Already Inferior to AI
In 2004, a team from Carnegie Mellon University began developing poker AI. They quickly managed to produce a program that would win in a three-card variant called Rhode-Island Hold ’em. Starting in 2006, the team entered its poker agents in computer competitions. In 2017, their software designed to play no-limit Texas hold ’em, Libratus, soundly beat four professional players. At the end of the contest, the four players lost a combined $1.8 million in simulated money.
In 2019, the AI called Pluribus, again developed by Carnegie Mellon University, in collaboration with Facebook’s AI Lab, defeated five poker pros in six-player no-limit Texas hold’em poker. The team of losing players included Chris Ferguson, winner of six World Series of Poker, and Darren Elias, the record-holder for most World Poker Tour titles. Pluribus beat all five players in the same game, and when each player went up against five copies of the software.
The developers state the reason why the software always comes out on top is that it conducts a real-time search to determine the proper strategy for a particular situation.
PokerBots Are Competing Against Themselves & Each Other
Poker AI develops by playing against itself. It does not study how other humans play, and its development process is that of a child. It utilizes something called a regret minimization training program, where it plays out hands in different ways when in similar scenarios, exploring what would have happened if it had folded instead of called, check instead of raised, and so on. It evaluates possibilities, noting the quality of each one, storing this info, and calling on the best course of action when in a similar circumstance.
MIT PokerBots is an annual tournament where MIT students have one month to program autonomous poker software. Competitors can code in whatever language they choose, and the competition features a prize pool of over $30,000. Naturally, this is low-level stuff to what we discussed above, but it may be an indicator of what lies ahead, a PokerBots World Tour.