In the study, published in Science, Tuomas Sandholm, professor of Computer Science, and Noam Brown, his student, details how his artificial intelligence managed to divide the game into computationally manageable parts and, with the game of its opponents, solve possible Weaknesses in their strategy during the competition. He did it with more decision points than atoms have in the universe. AI programs have defeated the best humans in chess and go, all challenge games, but in which both players know the exact state of the game at all times. Poker players, on the other hand, deal with hidden information: what cards their opponents have and if an opponent is bluffing. In a 20-day competition involving 120,000 hands at Rivers Casino in Pittsburgh, Libratus became the first machine to defeat the best human players in the heads-up no-limit Texas Hold’em. Libratus beat each of the players individually in the two-player game and collectively accumulated more than 2.1 million dollars in chips.
Win without analyzing poker faces
“The techniques in Libratus do not use expert knowledge or human data and are not specific to poker,” said Sandholm and Brown in the document. “Therefore, they apply to a large number of imperfect information sets.” Such hidden information is infinite in real-world strategic interactions, they noted, including business negotiation, cybersecurity, finance, prices and military applications. Libratus includes three main modules, the first of which calculates an abstraction of the game that is smaller and easier to solve than considering 10^161 (the number 1 followed by 161 zeros) possible decision points in the game. Next, it creates its own detailed strategy for the early rounds of Texas Hold’em and a rough strategy for later rounds. An example of these abstractions in poker is to group similar hands and treat them identically. “Intuitively, there’s little difference between a King-high ladder and a Queen-high colour,” Brown said. “Treating those hands as identical reduces the complexity of the game and, therefore, makes it easier from a computational point of view.” But in the final rounds of the game, a second module builds a new abstraction based on the state of the game. During the January competition, Libratus made this calculation using the Bridges computer from the Pittsburgh Supercomputing Center. Each time an opponent makes a move that is not in the abstraction, the module calculates a solution for this subgame that includes the movement of the opponent. Sandholm and Brown call this nested sub-game solution. The third module is designed to improve the strategy of the plan as the game progresses. Typically, Sandholm said, robots use machine learning to find mistakes in the opponent’s strategy and exploit them. Instead, Libratus’s self-execution module analyzes the size of opponents’ bets to detect potential holes in the strategy itself. Then, Libratus adds these missing decision branches, calculates strategies for them and adds them to the plan. In addition to beating human professionals, Libratus was evaluated against the best artificial intelligence in poker. These include Baby Tartanian8, a bot developed by Sandholm and Brown who won the 2016 Annual Computer Poker Contest held jointly with the Association for the Advancement of the Annual Artificial Intelligence Conference. The machines see a game as a tree. Simplifying, two branches come out of each node, which are the possible decisions or paths to take. For each of these branches fruit sprout, which are the possible reactions of the opponent. According to where the fruit has come from, so will two other branches emerge. Foliage and fruits compete for one goal: to reach the sunlight. Obviously, not all the branches are so leafy, nor all the fruits so compromising for them. Looking at the tree as a whole, from the bottom to the top, would give us a vision of the optimal path to reach the sun. But that takes time. For this reason, some branches can be clipped with their fruits, leaving it narrower. Neural networks are like experienced gardeners. They can learn which branches are typically the ones that reach higher or those that will bear the most fruit. The experience makes you score the branches according to whether they are more productive or leafy and thus help to make decisions about where to prune. So, what do you think about this? Simply share your views and thoughts in the comment section below.
Δ
