Pluribus Poker
'Pluribus achieved superhuman performance at multi-player poker, which is a recognized milestone in artificial intelligence and in game theory that has been open for decades,' said Tuomas Sandholm. Hand #1 - Hand #3400 13 elite poker pros were battling with an AI called Pluribus in a 6-max cash game (5H+1AI experiment). The poker pros were Jimmy Chou, S.
Let’s us look at the opening ranges of Pluribus when UTG/LJ, basically first to speak pre-flop; he can fold, limp, or bet… and actually never limps.
So the big surprise, if any, in the range is the bet sizes, going from 2BB to 3.5BB. Some of the other pros had a bit of variability there but not that much. Pluribus opens 18.6% from the LJ, which is a bit wider than what we’ve been used to with Snowie and others.
The only “real” surprise in actual hands is probably suited Kings, playing 100% with K8s, maybe with K7s (only 2 hands), and 2/3 of K6s. I find very odd that once each, AKo and JJ were not opened. Then more suited connectors than we usually see, with apparently a dislike for 87s. Obviously in order to open a bit wider, there are a few more offsuit broadways mixed in, like KT/JT at low frequency, and KJ at 55%.
Another surprise for me is there is not that much mixing up the hands, some, but very reasonable. Sizes however are all over the place.
Pluribus Poker Ranges – RFI LJ Let’s us look at the opening ranges of Pluribus when UTG/LJ, basically first to speak pre-flop; he can fold, limp, or bet and actually never limps. So the big surprise, if any, in the range is the bet sizes, going from 2BB to 3.5BB. Some of the other pros had a bit of variability there but not that much. GitHub - whatsdis/pluribus: Implementation of Pluribus by Noam Brown & Tuomas Sandholm, a Superhuman AI for 6-MAX No-Limit Holdem Poker Bot. Pluribus is a computer poker player using artificial intelligence built by Facebook's AI Lab and Carnegie Mellon University.Pluribus plays the poker variation no-limit Texas hold 'em and is 'the first bot to beat humans in a complex multiplayer competition'. According to the Pluribus creators, 'Developing a superhuman AI for multiplayer poker.
Displaying here ranges with some additional data on number of times Pluribus actually got the hand, percent open with it, and average, min, max bet sizes. Because the statistical significance of frequencies is limited (although not too bad for RFI), it is important to consider how many times Pluribus was in a given situation. For instance, he got AJo 17 times, and bet each time, we can be confident he opens at a high frequency with AJo. On the opposite with T8s, he got the hand 4 times at the LJ, and bet once, so we compute 25% but that could end up being 10% or 50%, all we know is sometimes he bets it, sometimes not.
So it is important to not take percentages for granted. Still the chart gives a very good idea of how Pluribus opens.
Pluribus Poker Program
Bet sizings are best looked as a distribution with percentage per sizing, in number of BBs. In general I would say sizing lower than what we’re used to, but with a lot of variation with sizes going up to 3.5BB. Either the CFR algorithms never converge on opening sizes so it stays more or less random or there is a real reason. Can’t really see what this achieves, again some of the pros did vary their sizings slightly so there must be a good reason to do it. What do you think ?
If you just want to download the hands to import them into your Poker Tracker, click here to download.
If you want to see the converter program source code, go here (you can also view all the converted hands in the out/ directory there).
background
Noam Brown and Tuomas Sandholm recently published a paper presenting Pluribus, the first no-limit non-heads-up poker AI that beat top professionals.
They released a download of some sample hands in the supplementary materials, but they were in a strange format where one hand looks like: STATE:42:r200fcffc/cr650cf/cr3000f:8c6h 7hJs AdJh 9s2h TdTs 5c3s/QdKc9d/Ac:-50 -200 -650 0 900 0:Budd MrWhite MrOrange Hattori MrBlue Pluribus
I wanted to get these hands into Poker Tracker to analyze them, but they were no good in the form they were in.
So I wrote a tool that converts the hands in the original format into hands in Poker Stars format. You can see it here. I've already found a couple bugs that I had to fix, so there are bound to be more. Please make a pull request or file an issue if you find anything wrong.
Cmu Poker Bot
I ran it on the entire dataset, and successfully imported it into Poker Tracker! You can download the dataset by clicking here.
instant analysis
To get the death of poker started, I did some basic analysis of Pluribus.
Pluribus Poker Bot Github
Here's Pluribus's graph:
Interestingly, it looks like Pluribus was actually negative in chips after the 10,000 hands. Either I made a mistake in the conversion code, and this graph is wrong (likely), or Pluribus was down in chips but up once adjusted with the variance-reduction (AIVAT) mentioned in the experiment (also very likely).
[7/19/19 edit: the graph is not wrong]
I did the most basic analysis I could think of - looking at Pluribus's preflop ranges. (Why did I not dive into anything more interesting than literally the most basic, uninteresting analysis possible? Honestly, it's because I'm pretty sure no one else has made this converter yet (at least not open source), so I'm trying to make use of all the clout equity that I have from being the first, so I'm releasing these hands as early as possible. In fact I barely got this preflop analysis together because it's almost 6 AM. I had to fix so many fucking formatting errors in the converter program. I don't even have Poker Stars, I just inferred the format from this wikipedia page that exists for some reason. I'm really glad that it does exist, though, and whoever added all those examples on that page is a legend, because you will not believe how impossible it is to find hand history file formats.)
Here's how often Pluribus will have each hand given that Pluribus voluntarily put money into the pot preflop:
Here's Pluribus' favorite hands to raise with preflop:
and its favorite hands to 3bet preflop:
Nothing too crazy here... looks like it 3bets AA, KK, AKs 100% of the time.
Poker Ai
Here's what it preflop raises as button or cutoff:
I'll hopefully do some more interesting analysis later, but I'm sure others with more Poker Tracker prowess than I can get a lot more out of this than I can, too.
Pluribus Poker Bot
Let me know if you find anything interesting @010010110000110