A new poker AI called ‘Pluribus’ has set a new milestone by beating some of the best players in the world at poker (specifically six-player No-Limit Hold’em), which could lead to a revolution. It’s very easy to see how this could completely disrupt online poker.
Thanks to deep learning and the tireless work of different AI researchers, AI has come along leaps and bounds in the past few years. Now, AI has set all kinds of new milestones, beating humans in various different games such as chess, Go, Starcraft and more. Now, a poker AI developed by Tuomas Sandholm (a researcher at Carnegie-Mellon University) has set new records. Sandholm had been working on developing software to beat humans at poker for years before finally reaching success with this new version. The original poker-playing software was named Cepheus, and now this version (capable of beating professional players) is named Libratus.
One of the most impressive parts of this new poker-playing AI is that it is able to thrive in a six-player game, as previously poker-playing AI could only participate in one on one games (“heads up”). Another major challenge in poker is that most of the information is unknown, unlike a game like chess for example. There is no way for the computer to know what cards its opponents are holding, and if someone bets, for example, there is no way for the computer (or anyone) to know for sure whether it is a bluff or not.
So how does this work? Like all AI, it was trained with a lot of data – the more data, the more the model will learn. In this case, the poker-AI played 5,000 games against five professional players, each of whom had won over $1 million playing poker. One surprising point, however, is that the cost to train the AI was actually obscenely low, only around $150 according to the information we received. That is amazingly cheap compared to many AI models that are estimated to cost millions in computing power.