Poker AI ‘Deepstack’ gazumps ‘Libratus’ as researchers claim victory over humans in no-limit hold’em clash

Poker AI ‘Deepstack’ gazumps ‘Libratus’ as researchers claim victory over humans in no-limit hold’em clash

Researchers from Canada and the Czech Republic have pissed on the Carnegie Mellon University Libratus parade by claiming that their AI, known as Deepstack, has already beaten human poker players in No-Limit Hold’em.

Poker AI ‘Deepstack’ gazumps ‘Libratus’ as researchers claim victory over humans in no-limit hold’em clashThe battle between the Carnegie Mellon University (CMU) Artificial Intelligence Libratus and four heads-up No-Limit Hold’em poker players has just developed a nasty case of blue balls after a group of researchers from Canada and the Czech Republic came forward to say it was nothing but yesterday’s news.

The brightest sparks from the University of Alberta, Edmonton, and a couple of Universities in Prague in the Czech Republic have joined forces to create DeepStack: Expert Level Artificial Intelligence in No-Limit Poker and they are claiming that it’s the first algorithm to beat humans in NLHE competition.

A non-peer review paper (found here) contends that Deepstack played 44,852 hands of poker, against 33 players, and came out on top, winning 492 mbb/g (Average winning rate over a number of games, measured in thousandths of big blinds). The researchers believe the pro poker player considers 50 mbb/g a ‘sizeable margin.’

The researchers wanted the players to compete in 3,000 games against Deepstack, and only 11 of them went the distance, including Phil Laak. I would love to tell you if Phil won or not, but only Bill Chen would be able to figure out the results.

I reached out to two of Deepstack’s opponent’s to gauge opinion.

“I played few hands, and gave up,” Said Luca Moschitta.”The software they used was that slow that it was making me very tilted and I realised I was playing poorly.”

Fintan Gavin faired a little better competing in 1,555 hands.

“I felt privileged to be offered the opportunity and found it to be a good overall experience with surprisingly decent software with almost zero glitches,” said Gavin. “The biggest challenge for me was completing the 3,000 hands within the time allotted.”

So how did this affect the outcome for Gavin?

“I crushed the bot in the first half of the match. But then in an attempt to speed up, I gambled a lot & lost all my later sessions.”

How did the bot play?

“The bot skill level was very mixed. While I believed, I could consistently beat the bot, if I wasn’t fresh & prepared and didn’t apply 100% concentration the bot beat me every time.”

How did Deepstack win?

Over to the paper:

It combines recursive reasoning to handle information asymmetry, decomposition to focus computation on the relevant decision, and a form of intuition about arbitrary poker situations that is automatically learned from self-play games using deep learning.

Forget about getting tilted by the software, Luca – I’m tilted by the use of language in the paper.

One line I did recognise in the maelstrom of complicated words was the researcher’s description of poker as the ‘quintessential game of imperfect information.’ And this is why No-Limit is the Holy Grail for AI researchers.

The researchers believe that Deepstack has the upper hand over Claudico, the CMU AI that lost against Doug Polk and Co last year because it operates differently than any other form of AI.

Back to that highly educated paper again.

DeepStack takes a fundamentally different approach. It continues to use the recursive reasoning of CFR to handle information asymmetry. However, it does not compute and store a complete strategy prior to play and so has no need for explicit abstraction. Instead it considers each particular situation as it arises during play, but not in isolation. It avoids reasoning about the entire remainder of the game by substituting the computation beyond a certain depth with a fast approximate estimate. This estimate can be thought of as DeepStack’s intuition: a gut feeling of the value of holding any possible private cards in any possible poker situation. Finally, DeepStack’s intuition, much like human intuition, needs to be trained. We train it with deep learning using examples generated from random poker situations. We show that DeepStack is theoretically sound, produces substantially less exploitable strategies than abstraction-based techniques, and is the first program to beat professional poker players at HUNL with a remarkable average win rate of over 450 mbb/g.

So that’s it then, there is no point in watching Libratus getting down and dirty with Jason Les and co because Deepstack has already proven that the future of online poker is fucked.

Or is it?

I would be willing to place a large wager vs. the bot over any amount of hands.” Said Gavin.

And in that sentence, Fintan Gavin described the future of online poker. The world where AI and humans co-exist across the whole spectrum of games both low and high stakes.