Pluribus Poker AI Dominates Human Players

The world of competitive poker has a new top dog in the face of Pluribus, a poker AI developed for nearly two decades and backed by scientists at both Carnegie Mellon University and Facebook’s AI division.

Pluribus is the result of these past 16 years of diligent research and the efforts of Carnegie Professor Tuomas Sandholm and PhD student Noam Brown, who works at Facebook AI. The researchers put the latest iteration of their AI to the test against two distinct poker professionals, to name Darren Ellis and Chris “Jesus” Ferguson.

Was the bar adequately high, you probably wonder? Absolutely, as Elias holds four World Series of Poker titles and Ferguson has won countless live tournaments. To avoid statistical inaccuracy, Pluribus played 5,000 hands against Elias and Ferguson each, winning with a fair margin over the opponents.

Let’s Talk Real Challenge: Pluribus Vs. Everyone

Both predecessors of the software, to name Claudico and Libratus, have been adept at taking on single opponents. No questions there. However, Pluribus has been a significant upgrade, allowing the software to handle multiple opponents at the same table.

Not that everyone played against Pluribus, but rather the AI’s ability to make decisions against multiple opponents was tested. Pluribus did well. The team behind the project had created a sort of a blueprint which allowed the AI to analyze the game and establish a basic behavior.

Even more, Pluribus was analyzing up to 6 million possible outcomes, going back as many as four plays for each individual player. This computing power, you would think, is what gave the AI it’s edge. Yes and no.

Pluribus’ Got Tricks Up Its Sleeve

According to players and researchers, the real advantage the AI had was its ability to avoid “limping” and use a strategy known as “donk betting”. These two terms are contradictory. Limping means that you involve yourself in a hand that you cannot win and that don’t want to raise.

This immediately signals players your inability or doubt to win. However, when Pluribus uses donk betting it basically “limps”, but then immediately raises on the next turn. People players interpret this as a buff or a sing of weakness, but Pluribus means business.

The upshot is that human players are met with an AI that completely randomizes its game to the point where it’s impossible to guess what’s going on next. This is precisely what poker pros have long hoped for themselves – to be unpredictable. Humans struggle with unpredictability, but computers can fake it very well, similar to a casino’s Random Number Generator (RNG).

Sending the Poker World in Disarray

One thing the Sandholm team won’t do is to release the source code of Pluribus. This way everyone can rest assured that the authenticity of poker events will be kept intact. It took the team 16 years to put together a working solution, and even though you could create it for $150, the hard skills and knowledge needed to develop it will take you a while.

Facebook and Carnegie Mellon University has decided to license their software to be used by military intelligence services and the video gaming industry. Yet, the implications of Pluribus can be much bigger when and if the company decide to license other industries.