The DeepStack algorithm was created by a group of Czech researchers working in collaboration with the team that first hashed out an algorithmic approach to Texas Hold’em poker, according to a report from Ars Technica. Like Libratus, it’s been able to beat professional opponents.
DeepStack was pitted against a field of 33 players sourced via the International Federation of Poker — although it’s worth noting that the competition wasn’t quite as fierce as it was for Libratus, as less prize money was on the table. Still, DeepStack ended up ahead against each of the 11 players who contested a full 3,000 game match, and was only down against two players overall at the end of play.
Texas Hold’em poker offers a different challenge to AI researchers than games like chess or Go, because each player only has access to a portion of the information they need to make a truly informed decision. For instance, the remaining cards in the deck in unpredictable because each player’s hand is private, unlike a game of chess where all the pieces in play are known from start to finish.
DeepStack avoids getting tied up in knots by this scenario by approaching every new action afresh. The algorithm is able to respond to any situation by performing a search based on the current state of the game, paired with a deep learning neural network that looks up the possible values of future hands.
Meanwhile, DeepStack chooses from a small selection of possible actions: Folding, calling, going all in, and making one of two or three different bets. This limited range of responses, coupled with the fact that the system doesn’t search forward to scope out every single future position, decreases the amount of states that need to be checked, which allows DeepStack to run on a GTX 1080. Conversely, Libratus was running on a system with a petaflop of computational hardware, and could perform calculations while its opponents slept at night.
DeepStack is certainly a promising project — and, like Libratus, it’s been engineered such that it’s capable of taking on any number of information-imperfect situations, not just Texas Hold’em poker. The end goal for its creators is to implement the algorithm in other applications where fast, accurate decision-making is key.
