Like RusyBrooks said, the C++ rand() function is not considered to be a great RNG (http://channel9.msdn.com/Events/Goin...idered-Harmful ). I have no idea how important this is for a software like pokersnowie, but obviously any serious poker room such as stars would use something more refined.

It is also quite possible that your learning algorithm may be biased if you use a poor RNG.

Not most people, indeed. But there have been claims about it (even ITT), and during a snowie challenge on a French website, a challenger made public his HH and claimed that snowie was cheating. By using snowie's preflop ranges I showed that snowie's rungood was very unlikely (~0.5%). If anyone wants to investigate further, or check my analysis, I can share the HH and my calculations. I have never considered this analysis to be a proof that snowie was cheating, but it obviously does legitimate doubts some players may have.

Therefore, even if this is not the main issue currently, I think it's good to see that they are making serious efforts towards more transparency.

I strongly back-up the other posters who have pointed out the massive deficiencies in using the C++ rand function (or for that matter any other simplistic RNG) for any task whereby good pseudo "randomness" is essential. Seeding it properly is another issue that needs serious consideration.

Personally, I try to avoid it like the plague, especially for anything whereby I am using stochastic methods e.g. MCMC, etc.

Agreed. Even Rand() within Excel sucks stained pants.

Hi RustyBrooks,

I fully agree with your comments. Finally it would mean playing on a third party client and server.

The features we added to export the deck and probs in order to give more control to the users were very important in my opinion, because the objective was to show that PokerSnowie is not taking any unfair advantage and it is not cheating while playing vs. the user.

About the random generator, I read some more comments and suggestions also from other users, like Babarberousse, TopPair2Pair and Oracle3001. Thanks for that. We are in fact aware that it's not the ideal RNG, but on the other hand we did not expect this to be an issue for the challenges purposes.

In particular we are using this one:
http://qt-project.org/doc/qt-4.8/qtglobal.html#qrand

Regarding the AI training, here a much more solid RNG is used.

Best Regards,

Roberto Gobbo
CEO - Snowie Games Ltd.

Hi Tackleberry,

you are right to wonder. In fact the answer is not easy, without entering into to many technical details.

Occasionally, in order to fix some issues in a very short term, we do like we wrote in the answer to David Sklansky: "for bet sizes that Snowie has never seen in training (like a 1% pot bet or a 100 times pot bet), algorithms have been implemented to handle these situations as good as possible."

Like you said, a full solution of an issue like this would need a new training to run and this would require some time. I admit that the short term solution in having some algorithms to handle these situations is suboptimal and it's more a patch than a real solution. When possible, we always try to avoid patches.

This point might seem conflicting with our statement that we don’t introduce expert knowledge in the AI, and in fact it partially is – apologies for that. But it should be clear that these patches are used only in very special and very uncommon situations, where we think it’s not worth to spend so much time, because they would not bring value to our customers. The 1% pot bet issue for example, is something which makes the AI technically exploitable, but on the other hand no one would play this strategy online and therefore fixing it was more to avoid this exploitation than to make it more useful for our customers. For this purpose, in the short term, we felt that a patch was enough.

In fact, because our CPU time is obviously limited, we need to choose on which aspects of PokerSnowie to focus. If we feel that how to play vs. a 1% pot bet is not crucial for the PokerSnowie goal, then we prefer to use our CPU for other improvements, like for example introducing other bet sizes. Obviously when you start a new training you can’t be sure if and when you will get the results.

Also, when I write “We fixed some of them”, I don’t always mean that the AI is ready. I mean that we fixed what we believe was causing the problem, but we still need to wait for the training to progress. Finally, in certain situation it’s not needed to start the training from scratch.
I hope this answer, at least at high level

Best Regards,

Roberto Gobbo
CEO - Snowie Games Ltd.

if you guys want some lolz heres a video using pokersnowie that makes little sense

http://www.pokersnowie.com/blog/2014...v#.UztY0PRDtv1

Why do you think it makes little sense?

Is that strategy video an April Fools joke?

Well one it doesn't bet the river because it can't balance it's range, but then it check/bombs in a spot where it can never balance it's check/bomb range?

It can't have a check range in the spot it bluffed shoved as it's facing a bet, and call ip. It does have a bet range in the spot it checked.

Pokersnowie does get there with some combos of the nut flush. It does bet most of them by the turn but checks some.

I think you need to watch the video again, you have the action sequence wrong.

And it checks those nut flushes on the turn and river?

I was sure it was CO vs SB, BB. But on looking again it's Co vs btn, SB. And it does bet all it's flushes on the turn when this is the case.

It check raises 12.28% of it's range. The only strong hand it has is a set of twos. Most of the time it has a pair below top pair, mainly a ten.

Exactly, and so the check raise jam line is complete nonsense on several levels.

lol wp snowie

PokerSnowie

Can you explain to subscribers (i.e. me), either here or within the software, the methodology used to train Team PokerSnowie members so that they were able to become winning cash game and tourney players within just a few weeks?

Thanks
Simon

Let's stop telling Snowie where it's leaks are, right now it has quite a few leaks, I know at least i'd prefer it that way, we don't need a bunch of bot type programs making their way onto the tables.

Seems like it'd only be a matter of time where someone creates a program outside of snowie using Snowie's strategy that sites can't detect. Not that i'd have a problem playing current snowie, an improved one, where ppl can use real-time would be disasterous.

Dear Elmo_Shedtax,

if you refer to Zeljko story published here:

http://www.pokersnowie.com/about/pok...timonials.html

as described, it took few months full time of hard work to reach a strong level.

There isn't a specific pre-defined methodology, as this field is all new and everything is evolving very fast. The first stage for a beginner who would like to learn poker by using our product, would be to play a lot of hands against PokerSnowie.

He could then review the hands, give a look to the blunders, the balance, learn the ranges in various spots, learn the PF strategy, drill down the areas of the game and see where he has have more weaknesses (btw, with our new 2.6 release coming out today, we have improved the filters, exactly to help this process).

As we all know, improving at poker requires a lot of hard work, discipline and we also recognize that more talented players could improve faster than others.

Best Regards,

Roberto Gobbo
CEO - Snowie Games Ltd.

Snowie has been trained for very short stacks up to very deep stacks (400 blinds). The neural networks try to avoid the biggest errors first. For very short stacks the size of possible errors is much smaller than for deep stacks, so it is possible that Snowie is in fact better for deep stacks than for short stacks.

The PokerSnowie Team

That's true Fontaine, and thanks for have tried it again.

Just to recap, originally you were min-betting into blown-up pots, like betting $1 into a $50 pot. Snowie didn't handle these situations well. An algorithm was implemented to react correctly to those mini-bets and the problem was fixed.

In general we believe it is very possible to win a 5000 hands challenge against Snowie, especially by pumping up the pot each time, the variance becomes extraordinary high, so that you have a good chance of being ahead after 5000 hands. A more reliable challenge would need many more hands.

The PokerSnowie Team

If Snowie fails to find decent solutions for a subgame where very strong solutions can be calculated in a matter of seconds using traditional techniques, what makes you assume it can find better solutions for cases where the complexity is dozens of magnitudes higher?

This is marketing talk at its best, unless you have data to back it up.

I don't think this objection is all that strong. The complexity of the task is relative to the resolution method. The short stack pre-flop games with simple decision trees admit of precise analytic solutions that can be computed easily. But neural-network training algorithms can sometimes rapidly find solutions to problems that don't admit of analytic solutions at all and, on the other hand, can be stumped by problems that can easily be solved analytically. It's more akin to an empirical trial and error method where we have an idea of what a good result looks like but we don't understand at all (initially) why some method or other works best. Of course, when an efficient method is found -- semi-empirically -- we can investigate why it works and, hopefully, derive general principles.

That said, it's true that performance in cases where we know the true GTO solution provide an ideal benchmark to the effectiveness of the training algorithm. I'll give you that.

By "traditional techniques" i was actually referring to iterative algorithms like Fictitious Play or CFRM, not solving analytically. These algorithms actually very roughly resemble the iterative training of a neural-network.

Your example regarding "simple" problems stumping neural nets while much harder problems can be solved very well is misleading. This typically refers to problems with entirely different structure, which is not the case here imo.

Can you give an example of a problem where neural nets fail to find good solutions to very simple instances but succeed in solving huge instances of the same general structure?

(On Edit: I regrettably confused the push/fold game with the raise/shove game below. That seriously undermines my argument, but I'll leave it since it raises an issue regarding training algorithms and the indifference principle for GTO mixed strategies.)

That's a good question.

In the case of the push/fold game, it was brought up that the ranges were wrong. How about the frequencies for the bet-call action or the combined (open-fold + bet-fold) frequency? It might be more difficult for a neural network training algorithm to home in on all the correct GTO ranges in this sort of cases because the equilibrium condition dictates that open-folded hands have the exact same EV as raise-folded hands at that decision point. (At equilibrium, the BB shove frequency must be such that the SB is indifferent between the two actions with hands that never see a showdown). The very slight differences in EV stems only from card removal effects on the BB shoving frequency. But those effects can either be computed analytically or discovered by a neural network training method that is focused on finding them using huge samples that would make those small effects statistically significant in between random choices of ranges. Pokersnowie's training algorithm understandably isn't optimized for defining ranges that have so very little effects on overall EV.

This failure for Snowie to define the proper open-fold and raise-fold ranges then will translate in a corresponding failure to home in on the correct individual frequencies for those two action. (It ought to find the correct calling frequency, though, since the larger EV of the hands he can call a shove with dictate the proper action).

More complex decision trees that don't include such indifference points will possibly more easily be solved -- or will be solved modulo the detailed partition of ranges between the (roughly) indifferent decisions when there are some such points.

I agree that indifference will likely not play the same role deep-stacked that it does in the raise/fold game. (And on a side note, the raise/fold game also gives the other iterative algorithms i mentioned earlier a fairly hard time.)

The general indifference problem is still present whenever we have pure bluffs as part of a range though. So at least on the river this can be expected to be an extremely common situation deep-stacked. Snowie needs to somehow figure out the correct bluffing ratio based on the tiny EV differences caused by removal. If it fails to do this for simple Preflop games i would not expect Snowie to solve the same general problem Postflop.

There are plenty of other factors that make the deep-stacked games much, much harder to solve generally, so we really have absolutely no idea if Snowie can solve them better. Theoretically possible? I guess. But absent any evidence to support it, that's going to be very tough to sell.