Heads-Up Texas Hold’em Algorithms Now Hard To Beat
There is a new artificial intelligence algorithm to compete with in the previously weak space that includes games of imperfect knowledge such as poker. It goes by the name of DeepStack, and judging by the results of some heads-up games of poker against a panel of experts it looks to be fierce competition that will require the humans to tweak their own strategies if they want to maintain their bankrolls. At the moment, the research has not yet readied itself to compete against full tables of professionals, but as with most everything in the technology revolution that day seems closer than ever.
Games of imperfect knowledge or quite a different obstacle than those where the entire state of the system is known such as in a chess game where the position of every piece can be tracked. In poker every player has their own knowledge which differs from the other players since their hole cards are personal. Limit Texas Hold’em is an ideal application for these newer algorithms such as DeepStack since it allows only a small amount of variation in the action to be chosen: fold, call, go all-in, or make only two or three different bets. No Limit Texas Hold’em can be treated in the same way such that only a limited number of bets are possible, and that is a simplification which allows the artificial intelligence to play equally as strong in such instances.
Nash equilibria and counterfactual regret minimization are the hot words of the hour and together make up most of the concepts used for strategy. The fascinating part is that while just a year or so ago it would take a cluster quite some time to reach a point where the strategy problem was ‘weakly-solved,’ an Nvidia GeForce GTX 1080 video card can now make the computation in about 5 seconds using the latest innovations. The value of possible future hands was also an important consideration which required a couple of deep learning neural networks in order to provide that information to the system.
The end result was that some astute observations can now be considered highly profitable such as holding a small pair when playing from the dealer position. That ability to act last is one of the greatest advantages that any player gets, but many such standards of action have already been known for quite some time and written about in the myriad of literature that exists for Texas Hold’em in all of its forms. One the more interesting revelations is that the highest level players did not make themselves available for testing. This is probably greatly due to the amount of compensation offered which was not enough to lure that level of player, but it is also highly likely that poker players are beginning to understand that the maturity level of opponent algorithms is now reaching a point where they will be hard to beat.
This could mean the effective end to online gaming since it will be hard to detect anyone sitting on the other side of a computer screen with his decisions being computed on a graphics card, but the future still remains to be seen and anti-cheating measures will have to up their game as well.