In case you missed it, April 24 – May 8, 2015 saw four of the best Heads Up No Limit players in the world face off against advanced poker bot 'Claudico', developed by scientists at Carnegie Mellon University (CMU).
On the human team were Doug Polk (WCGRider on PokerStars), Bjorn Li, Dong Kim and Jason Les.
The 'Brains vs Artificial Intelligence Poker Challenge' was the first HUNL competition between man and machine - thus far computers have only mastered Fixed Limit.
The stakes were $50/$100 and players started each hand with 200 big blinds. Over an 80k hand sample the humans beat the bot with a 9bb/100 winrate. Although statistically the match was concluded a tie, mankind can rest assured that S̶k̶y̶n̶e̶t̶ artificial intelligence isn't ready to take over the world just yet.
Bjorn Li beat the bot by the highest margin, finishing +$529,033. He joined us for a post-match interview.
Rakeback.com: Congrats! Although the match ended a statistical tie - the researchers not being able to say with over 95% confidence that the human players’ chip lead was scientifically reliable – you beat the bot for by far the highest amount of the group. What do you put your success down to?
Bjorn Li: I guess I would say it was partly run good and partly doing the right things. I thought the bot was overfolding in a lot of spots so I was trying to exploit it as much as I could knowing it is a static GTO bot that doesn't readjust. Also to win the amount I won I obviously still had to run really good situationally, for example I tend to get river folds more when I had a bluff and river calls more when I had a value hand. So overall there is still going to be a lot of variance at play here.
RB: You won most of your chips prior to the halfway point of the match, and there was talk of Claudico playing better in the second half having made adjustments. Did you feel you were up against a tougher opponent towards the end, and if so do you think it could have eventually turned the tide?
BL: It made some improvements for sure, although it wasn't nearly enough to be able to beat us by the end of the competition. For example, it cut out on really big mistakes in terms of bb loss per hand by removing its largest sizing (all-in) as it will not longer do that with the wrong hand selection. However, there are a few other areas that is still leaking a lot on.
Match view via Twitch. Li mixing it up, raising 6x pre with the old 3♠2♥
RB: WCGRider noted that at some points Claudico would underbet or overbet the pot by a huge margin, e.g. betting $19,250 into a $700 pot, something ‘a person would never do’. The bot’s team were perplexed and couldn’t analyze enough of Claudico’s set strategy - two terabytes of data – to work out why. Any idea as to why a bot would take those lines?
BL: Related to the last question. In the first half of the match, Claudico has the all-in sizing on rivers which in itself can be a fine strategy. The only problem was that it was not doing it with the right ranges and specifically does not use blocker considerations. Using a large range of sizings in every given particular spot is definitely the way to go as far as a GTO solution/Nash Equilibrium goes, so I definitely wouldn't say its perplexing in my eyes. Its just a matter of whether its implemented well.
RB: Overall the human team had a 9bb/100 winrate over the sample of 80k mirrored hands. WCGRider said in a post-match interview Claudico was a very strong opponent but he doesn’t really feel it was tie based on those numbers. Do you feel the same? Researchers commented that some of the bot’s weaknesses weren’t captured in the statistics.
BL: Fact is that the human team won at a 9bb/100 winrate over a 92% confidence interval. So if you asked me whether I feel its a "tie" I would say that there's a 8% chance humans are actually true losers in the match, but I think there's a good chance that % is actually lower than that. I definitely think humans also weren't playing under optimal conditions as we had to deal with fatigue and boredom as well as having to deal with publicly exposed hole cards.
RB: WCGRider said he believes Claudico would beat stakes up to HU NL400. Your thoughts?
It will beat most regs at 400-600HUNL. But 1knl and higher is probably harder.
WCGRider's post-match interview after the Brains vs AI competition
RB: Do you feel playing the most advanced poker bot in the world has added value to your own game, improving your play going forward? If so, how?
BL: I would say that almost everything good that the bot is doing are not things that I don't already know exist strategically, and are already most things I implement in my own game. For example using multiple sizings on rivers is already considered a gold standard in today's HUNL games, although this might not be immediately obvious to a lot of people. It also does this 1/10 minlead on rivers after check calling 2 streets that I think can actually work out pretty well in real games as well.
RB: You said in a post-match interview that we know AI is ‘going to overtake us one day’. How close do you think we are to reaching that day for No Limit Hold’em?
BL: I think it can be anywhere in between 1-6 years. Its only a matter of time that it will because in HUNL there is a GTO solution/Nash Equilibrium and when a bot is eventually able to play perfect Nash, it will be unbeatable (i.e it has no exploitable leaks).
RB: You’ve played WCGRider quite a few times on PokerStars, and he said in interviews he’s sure in you’ll have ‘some battles down the line’ in future. He also ranked himself as the #1 HUNL player online. Any thoughts on your edge over him, and when you’ll face off again?
BL: For now he probably still has an edge over me probably at a range of 1-4bbs/100, although note the multiple strategic usages of the word "probably." We'll definitely face off again at some point, but I feel like both of us have got other better and more profitable things lined up for now.
Thanks again Bjorn!
Up to 30% Rakeback + €1000 Bonus
Up to 70% Rakeback + None Bonus
Up to 22.2% VIP Rakeback + $500 Bonus
Up to 30% VIP Rakeback + $600 Bonus
Up to 30% Rakeback + $1000 Bonus
Up to 27% Rakeback + $1000 Bonus
Estimate How Much Rake you are paying to your poker room, and see how much you could earn back!