In 2017, we highlighted the first major example of an artificial intelligence program beating human poker experts. in particular, a program designed by Carnegie Mellon (dubbed “Libratus”) was deployed at Pittsburgh’s Rivers Casino, where it pitted four professional poker players into heads-up poker over a span of 20 days. After 120,000 hands, Libratus had amassed $1,766,250 more than all four pros combined, indicating a real advantage over the pros. As we told at the time, computer science professor Tuomas Sandholm concluded that the “ability of AI to reason strategically with imperfect information” had surpassed that of “the best humans”.
However, it would turn out that Libratus’ victory was just a milestone on the road to an even more impressive feat. Two years later, a follow-up program – this one called Pluribus – would deliver even greater achievements in poker. As explained in a article about More at ScienceDaily.combeat this second program six poker professionals decide over the course of 5,000 hands. In a separate experiment, it also hired 13 poker professionals who each scooped $1 million or more in in-game revenue –– playing five at a time and coming out on top after 10,000 hands.
The AI programs and the experiments they participated in were similarly set up in some ways. Both programs are designed not only to understand and play poker, but specifically to participate in no-limit Texas Hold’em. As many readers will know, Hold’em in general is by far the most popular style of poker –– but the distinction between “no-limit” is important. Poker.org Defines No-Limit Texas Hold’em as a variant of the game in which “players can bet at any time in the hand up to the size of their pile as they wish.” This is in contrast to other types of Hold’em which have different limits on how much can be wagered at one time; that’s why we’re aware of the concept of “going all in,” as seen in movies, televised poker matches, and so on. And from an AI poker performance standpoint, it makes Libratus’ wins and Pluribus all the more impressive. Fewer restrictions mean more possibilities and more variety; no-limit forces the programs to intuition the best games consider putting pressure on opponents – as opposed to simply making decisions they think are mathematically advisable.
Where the Pluribus feat stands out, however, is the number of players involved – not just in total, but at any given time. As noted, Libratus defeated poker pros in a heads up no-limit Hold’em construction. In poker, that essentially means playing one-on-one. The program consistently outsmarted four poker professionals, but did so by beating them one by one. This makes for a slightly simplified game play, while Pluribus faced the added challenge of taking on multiple opponents at once. And in multiple test scenarios, the newer AI proved to be able to outperform all of its adversaries simultaneously –– process much more information and weigh strategies against each competitor at the same time. It was not only more challenging, but also a more accurate test of whether an AI could beat top human players at a standard no-limit Hold’em table or tournament.
What this means in the long run remains to be seen. In computer science and artificial intelligence research, there is something to be said for the idea of achievement for the sake of achievement; scientists wondered if an AI could master poker more effectively than a human professional, and proved that in fact it could. On the other hand, Pluribus’ achievement can also pave the way for further innovation. In a article about Pluribus on DigitalTrends.comit was noted that co-creator Noam Brown believes the bot’s “ability to handle multiplayer, hidden information, and myriad possible outcomes” could lead to real-world applications “for the benefit of humanity.”
Only time will tell what those applications might be. But we might one day look back to see that the victories of Libratus and Pluribus in poker paved the way for significant AI use.