facebook rss twitter

Nvidia GPU powered DeepStack AI beats poker professionals

by Mark Tyson on 4 April 2017, 12:31


Quick Link: HEXUS.net/qadfy2

Add to My Vault: x

Nvidia boasts of passing a new milestone in gaming AI. Scientists from the University of Alberta's Computer Poker Research Group, with help from researchers at Charles University in Prague and the Czech Technical University, managed to create a GPU-trained AI that can defeat poker professionals in heads-up No Limit Texas Hold'em.

Poker is an interesting game for an AI to succeed in. Unlike previous AI gaming milestones where players, including the computer, have 'perfect information' - poker players are challenged by many unknowns. In Chess, Draughts, or Go for example, all the pieces can be seen on the board, the 'imperfect information' aspect of poker comes from the many unknown and unrevealed cards in a game. Add to this uncertainty the intuition and strategy honed by poker masters over years of play, and its difficult to train an AI to win.

The success of the GPU-trained DeepStack AI has repercussions outside of gaming. In the real-world there are very many problems to solve where imperfect information is a feature. DeepStack's training and success in using its training to win shows that similar GPU AI training could be used for complex problems in business, industry, science and so on.

Interestingly, now that it has been trained, DeepStack can "run on a gaming laptop" (with a GTX 1080 GPU). Meanwhile a rival No Limit Texas Hold'em playing supercomputer called Libratus (which uses 'a thousand CPUs', cost nearly $10 million, and packs 274 terabytes of RAM), is about half as fast as DeepStack in gaming decisions, say the CPRG researchers.

If you are interested in hearing or reading more about DeepStack, the Nvidia blog features an interview with Michael Bowling, a professor in the University of Alberta's Faculty of Science (Bowling is centre in the above group picture). For further reading you can head on over to the CPRG home pages.

HEXUS Forums :: 3 Comments

Login with Forum Account

Don't have an account? Register today!
RIP online poker?
From a game theory point of view, it doesn't actually take that much skill to beat certain professionals, who look to exploit the low-hanging fruit in the form of awful players, and you can in turn exploit them. People are also bad at adapting in a truly random fashion, something AI has always held as an advantage over them, so it's not a particularly big deal and AI has been beating professional players for a while. This is of interest because it's no-limit as opposed to limit, in other words the range of choices isn't limited to ‘bet, check, raise, fold, call’, but when betting or raising, includes amounts as well, effectively turning the game from discrete to continuous. Under these conditions human instinct initially does better than AI, but once AI catches up, it gives it more scope for crushingness.

What's interesting to me is that a self-learning AI tends towards strategies previously considered ‘outdated’ by high-level human players, and they are now attempting to incorporate them into their respective games.

Online poker will never fully die, the less secure sites may see their traffic drop when people are effectively unable to win without programming their own bots. Thus far the industry hasn't been sufficiently incentivized to keep bots and other forms of cheating (card-sharing via internet messenger) minimized, and it could well be by the time they have done so, their supply of fresh blood will mostly drain out, meaning we'll see a ton of consolidation.
Not sure how this is anything new given you can play heads up with nash equilibrium strategy…