Inside PokerSnowie's brain reveals the work of the Snowie AI Team. It explores first hand how the brain of PokerSnowie evolves and learns advanced strategic concepts, on its own.

PokerSnowie's ultimate aim is to produce the perfectly balanced game, find the ultimate un-exploitable equilibrium for all No Limit Hold'em configurations. Join us on this fascinating journey, which is just starting, into the future of poker.

The PokerSnowie Video Quiz series aims to answer the question: "what is the right play?" It is a set of poker coaching videos presented by French professional poker player Sharp. In each video, Sharp sets up an interesting hand in the "Scenarios" tool of PokerSnowie and explains how to analyse the situation and learn from PokerSnowie's advice, based on the Game Theory Optimal model.

1,5 million hands played - PokerSnowie still crushing it!

The Challenge PokerSnowie contest is still going strongly, approaching 1.5 million hands played. 

So how is PokerSnowie doing against its many opponents?

 

The overall precise figures stand at 1,459,325 hands played, an average error rate of 16.92 and a global winrate for PokerSnowie of 22.18 bb/100

As usual, we removed anything under 100 hands played, and any play with an error rate above 25.

Results:

  • 1,232,481 hands played
  • An average error rate of 14.14 (advanced).
  • PokerSnowie made a profit of 15.53 bb/100

How did the results by Error Rate evole? As per last update, there is still a strong correlation between win rate and error rate. However, with the exception of the Extra-terrestrial level (which is still not statistically significant), all level of players are now losing against PokerSnowie in aggregate. 

PokerSnowie's win rate by Error Rate level
  Hands played Error rate PokerSnowie's win rate
(in bb/100)
Extra Terrestrial
(0-4)
1,825 3.55 -38.04
World Class
(4-8)
115,515 6.54 +0.34
Expert
(8-12)
341,947 9.88 +5.40
Advanced
(12-16)
325,446 13.80 +15.29
Intermediate
(16-20)
266,289

17.48

+21.12
Beginner
(20-25)
174,769 21.58 +38.43

 

How did the results by game type evolve? (considering only $0.5/$1 games)

Challenge PokerSnowie results by game type
  Hands Played PokerSnowie's win rate
(in bb/100)
Head's Up 314,818 +23.1
6-Max 367,916 +13.9
Full Ring 108,298 +21.4

 

And finally, there are now 47 players who individually played in excess of 5000 hands against PokerSnowie. They played at an advanced level of 14.32, over a total of 387,948 hands. And PokerSnowie is still winning 11.05bb/100 overall.

The list of individual winners is however also growing, with 18 players ahead of pokerSnowie. User "Fontaine" managed the biggest result with a win rate of 36.29bb/100 over 6042 hands. A special mention as well to user "Matthew" who recorded a win rate of 3.86bb/100 over 11,461 hands, playing at a world class level of 7.46!

We thought it might be interesting as well to check the overall results by error level for players having played more then 5000 hands against PokerSnowie: 

PokerSnowie's win rate by Error Rate level
  Hands played Error rate PokerSnowie's win rate
(in bb/100)
Extra Terrestrial
(0-4)
0 NA NA
World Class
(4-8)
46,656 6.67 -1.05
Expert
(8-12)
105,489 10.09 +10.38
Advanced
(12-16)
114,234 13.85 +11.08
Intermediate
(16-20)
81,352

17.37

+19.34
Beginner
(20-25)
17,963 21.27 +20.33
Beginner
(Above 20)
40,748 23.04 +8.49

Once again, we can notice a correlation between the win rate and the level of play - with the exception of Fontaine's win, where he played at an error rate above 25.

 

We have been overwhelmed by the amount of positive feedback - as well as your constructive criticism - received on the Challenge PokerSnowie feature and how playing against PokerSnowie and analysing the results in PokerCoach is helping your game. Many thanks for all the feedback, and please keep sending us your comments, requests, ideas and other thoughts - they all help towards continuing to make PokerSnowie the best possible Poker Coach.  

Happy New Year to all!

Comments

Among winning players, am I the only one with an error rate of only 7.11? If so, can I get a shout-out in this blog entry as well? I'd love the bragging rights among my friends! Thanks!
Hi Jay, amongst winning players (5000 hands +), you have the third best error rate. The best is 5.23, for just under 6000 hands and win rate of 0.02bb/100; Second player has an error rate of 5.56, also for 6000 hands and a win rate of 0.28bb/100. And finally, you come third with a very strong win rate of 2.98 over 5000 hands - and error rate of 7.11 indeed! Do you think you can improve this over the next 5000?
Official winner's mention: Carlos played 5803 hands, with an expert error rate of 8.86 - and a winrate of 6.31 bb/100. Well done!
Just done 5k hand on the Challenge (Its pretty fun!) with an error rate of 6.96 and 6.83 bb/100 (if I'm not wrong). Could you please double check it just to know if I miss something? Thank you!
can you send us an email to support@pokersnowie.com, so we can look up your account and confirm? Thanks .
Have you been improving PokerSnowie on a regular basis since the challenge started? It seems reading various forums and threads that it's exploitable, more so in 6-max cash and less so in HU cash, but it's still far from optimal. The challenge has now become very important for PokerSnowie's improvement. Do you have a roadmap of the improvements that you will make to PokerSnowie? As you continue to improve PokerSnowie, you will need to reset the statistics, as they will not be indicative of the improvements to PokerSnowie. Furthermore, it's preferable if you split HU cash and 6-max cash statistics, as PokerSnowie is always going to be stronger at HU cash.
Hi Chrono. Thanks for your comments. The challenge function is a tool for players to test their game against PokerSnowie and then analyse it in PokerCoach. We do not use the hands from challenge in any way to improve or develop the AI. This is strictly done through the neural network - anything else would not enable it to follow the GTO approach! As for PokerSnowie's exploitability - we have seen just as many comments about it being unbeatable as we have seen the opposite. The overall results, which we detail on a regular basis on this blog (including HU, 6-max and full ring details), clearly show, so far, that PokerSnowie is winning a lot more then it loses. With regard to the roadmap, we intend to release a new PokerSnowie brain in the next few weeks, based on the last 6 months of training on the neural networks. We will obviously announce it officially prior to its release. Thank you, as ever, for your feedback.
Yes, but you can modify, remove, or add parameters for the training method for PokerSnowie's neural network. Studying the strongest hands of the challenge will allow you to improve the parameters for the training method. I assume that the people behind PokerSnowie have studied "The Mathematics of Poker" by Chen and Ankenman. However, there are two books that go significantly beyond it, called "Expert Heads Up No Limit Hold'Em, Volume 1: Optimal and Exploitative Strategies" and "Expert Heads Up No Limit Hold'Em, Volume 2: Strategies for Multiple Streets" by Will Tipton. The second book is going to be released towards March to May of this year, earlier in the US and later in the rest of the world. Here are the links to publisher's store for the two books: 1) http://www.dandbpoker.com/product/expert-heads-up-no-limit-holdem-volume-1 2) http://www.dandbpoker.com/product/expert-heads-up-no-limit-holdem-volume-2. Errata, discussions, and various utilities are inside the first link, and they will probably also come for the second link. Some more discussions on the first book are available in this thread: http://forumserver.twoplustwo.com/33/books-publications/expert-heads-up-no-limit-holdem-v-1-optimal-exploitive-strategies-will-tipton-1281938/, while a similar thread in that forum will most certainly also appear when the second book is released. I assume that the team is based in Europe, and therefore they should import the second volume directly from the US, as the rest of the world is going to have it one month later. I don't know if translations have been made for the first volume, however I suggest to study the first book's original English version if the publisher has made translations as translations are often bad and full of mistakes, especially for technically advanced books such as these. I suggest that the team deeply study these two books, as they will help understand the hands gathered from the challenge and PokerSnowie's weaknesses. These two books will allow the team to understand what improvements need to be made to the parameters for the training method for PokerSnowie's neural network. Midstakes and highstakes NL HU cash regulars will find these two books useful, and given that the skill of the developers behind PokerSnowie is going to directly affect the training method for PokerSnowie's neural network, it is a requirement that the team attempts any reasonable method for improving their understanding of the game, so that the parameters for the training method for PokerSnowie's neural network can be significantly improved. Otherwise, PokerSnowie's skill level is probably going to remain relatively fixed and very exploitable for years to come. If years of a training method similar to the current one have made PokerSnowie reach this current level of skill, then months and years of the current training method are probably not going to improve PokerSnowie significantly. After all, PokerSnowie being almost breakeven against players with an error rate of 4-8 at a statistically significant sample of more than 100k hands implies that PokerSnowie is still quite far from being optimal. Most of those players have the skill of average midstakes regulars. The best 6-max midstakes regulars would easily be able to have a significant winrate against PokerSnowie at 6-max and at more than 100k hands. The best measure of PokerSnowie's skill is at HU, and even at HU the best midstakes HU regulars would be able to beat it, but probably at a relatively smaller winrate. The team has done an impressive job until now, as many people thought that an AI of PokerSnowie's level of skill would currently be impossible for NL HU and 6-max, but the team also have a difficult and long road ahead. The team will need to improve their knowledge of poker, improve their knowledge of computer programming and computer science, have greater computational resources, and employ even more intelligent parameters for the training method for PokerSnowie's neural network in order to make PokerSnowie achieve a higher level of skill. The team should also make PokerSnowie be able to bet 3/4 pot and other bet sizes, as bet sizing is a significant part of PokerSnowie's weakness. Finally, PokerSnowie is much stronger HU, so it doesn't make sense to combine the HU and 6-max challenge statistics.
What I mean is that you should provide winrate by error level for HU, 6-max, and fullring, all three separately. You should look into using cloud computing in order to have a lower budget: 1) http://aws.amazon.com/ec2/spot-and-science/ 2) http://www.cyclecomputing.com/. Amazon likes to use CPUs and doesn't really like GPUs, so for perhaps an even lower budget, if the training method can be made to be sufficiently parallel, you could just use colocation and buy Nvidia or AMD normal GPUs if working with single-precision calculation only is ok, or the much more expensive Nvidia or AMD compute cards if double-precision is needed. The best programming language to exploit parallelization is Haskell. If you want to use Haskell for GPU compute, here are three resources: 1) http://learnyouahaskell.com/chapters 2) "Purely Functional Algorithm Design" by Richard Bird 3) "Parallel and Concurrent Programming in Haskell" by Simon Marlow.
A probably even smarter colocation idea would be to colocate a mixture of AMD Kaveri APUs (they combine CPU and GPU through an unified memory architecture, which allows for algorithms that are slightly less parallel than ideal to be used) and AMD GPUs (and / or compute cards). The AMD Kaveri APUs and AMD GPUs and / or Compute Cards are amplified and connected through the HSAIL finalizer: http://hsafoundation.com/hsa-developer-tools/
2) should be "Pearls of Functional Algorithm Design" by Richard Bird.
Hi, Have you thought about adding a stacksize option to Snowie Challenge? I play with 20-40 BB in my games for the most part so training with 100BB is not all that useful for me :/ Besides that I like the new challenge feature, thanks a lot for adding it!
Thanks Kevin; Yes, we are planning to have a short stack and deepstack option added.
Looking at the Pokersnowie calling range in the SB vs. a Pot BTN open the only hand in it's range to call is KTs which it flats 23% of the time. This can't be correct it should flat a wider range to be harder to read or it should not have a flatting range. If the pot is opened on the BTN for pot with TT and Pokersnowie calls in the SB if the flop comes T K K the best play is to give up unless we improve to quads without having to pay to get there.
Hi Andy, this is a specific case where PokerSnowie's range is very narrow, which indeed leads to potential problems which the opponent could exploit.
Slight error in last post. Can't make qauds with TT on TKK when snowie has KT.