So raising here has an EV of -$3.64?
Didn't know you could rebuy while in a hand these days. Confirmed: Poker is dead, everyone is Snowie.

yeah, think abput your own range.
well, i guess, you are right.

No, that is not the purpose of GTO. GTO is GTO regardless of your opponent's range, and it does not aim to exploit except in the very most technical sense of the word (GTO maximally exploits another GTO strategy, but not necessarily anything else)

Maybe you'd like to purchase one of my fusion reactors? It's not technically fusion but I'm using this method to build a fusion reactor and this is as close as I can get. You in?

OK actually it's a wood burning stove. After you buy it you can have a good time telling everyone how nice it is being able to see how fusion works from a first hand perspective.

Rusty, a couple of questions -

- At the moment it's obviously quite useless, but given the rate of expansion of knowledge and technology, how do you see it's prospects doing over the next year/3 years/10 years?

- Maybe a perfect game can't be produced, but how perfect does it need to be until it is effectively unbeatable? I'm pulling numbers out of my ass here and not sure this even makes sense, but say it progressed to a point where it had the game 20% solved, could it approximate the other 80% to a satisfactory point? Would it be possible to solve one or two streets (ie preflop then flop) and then turn the decisions over to the player?

That's most likely in bb's. So raising has an EV of -$.364 (according to them).

I don't really know, and I don't think anyone else did.

The thing that grinds my gears here is, look at the available literature on GTO solved games. Like, really look at it. The kinds of games that have been solved are *so* completely simple compared to NLHE. Miles away. In fact I think the distance is adequately summarized by comparing wood stoves to fusion. The "near GTO" claim for HU NLHE is bad enough, for multiplayer games it's just laughable.

I'm not a specialist by any mean but yet, I think there's so much misunderstanding going on here (and basically on almost any thread about game theory for poker).
RustyBrooks raises a very elementary but overlooked issue. Game theory studies "models of games". For some (classes of) models, we are currently able to prove the existence of GTO strategies. For very few of them, we are currently able to exactly compute them.
First and foremost, game theory may or may not help you prove the existence / approximate / exactly compute GTO strategies for models that more or less accurately describe real poker. But it's just models of it. Not the true game we are actually playing.
And yes, at this point, we can exactly compute GTO strategies for models that describe poker so roughly that they are very informative regarding what a GTO strategy would be for something that even starts looking like real poker (at least for multi-way, NL games).

This being said, it may not be necessary to be able to EXACTLY find GTO strategies. Find approximate solutions can be of great interest. From a theoretical point of view, though, it can only be satisfying if you are able to quantify "how far" from the optimum your approximation is. And this is not a trivial issue. Even defining what a relevant distance (to measure this gap) is non-trivial. What we seem to be able to do now is, given a fixed strategy, to compute the optimal strategy against it. Then, we can evaluate how it performs versus this optimal counter-strategy. If it break evens, it's GTO, if not, it will necessarily be losing and the amount by which it is losing can be a measure of "how fare we are". But it is a pretty bad measure, though. I don't want to get too technical, but GTO strategies are solutions of optimization problems (ie minimzing/maximizing a function). If the function you are optimizing does not have some desirable (regularity) properties, the EV gap tells you nothing about how far the strategy itself is from GTO. You could have a huge EV gap with a strategy that is in fact very close to optimal (if the function varies very fast around its optimum). You could also have the opposite: find a solution that has a very tiny EV gap, but that actually is very very far from the true optimum.

All this to say that there is not necessarily anything wrong with pokersnowie. But to evaluate that, they should provide more details about a) what model they are studying, b) the function they are optimizing (and some of its properties), the algorithm they use to approximate it, or at least some analysis of how their approximation relates to the true optimum of their function. I checked their website. I found none of that.
And no, even if they made public the full characterization of their so-called GTO strategy, it would not be enough. We would not be able to say anything about it, except that it would not be truly optimal (but no one should be expecting that anyway). We need the solution and how they reached it. Else, we just can believe them. Or not. Because it's so full of imprecision that I doubt it can be scientifically sound.

The problem at this spot is that villain's cbetting range is empty (Snowie advice is to check with the whole range). When 1st player cbets, Snowie advice for Hero (2nd player) is to fold QQ, call KK+ and AQcc. Snowie assumption is that 3rd player's range is very strong and is EXACTLY QQ (I believe this is flaw of Snowie putting 3rd player only to one particular hand with 100% frequency). That means first player cbets only QQ+ and better (but under assumptions of Snowie it should check the whole range to maximize EV).

Having said all of that, advices given by Snowie not follow GTO even closely for multiplayer games.

i put this spot into snowie at 1/2 and it telles me to raise 1%o of the time and fold 99% booth moves are 0 EV and calling is slightly -EV, 0.-8 to be exact.

So its almost all the same and quite different to what you are posting.

This could be due to rake (snowie is aware of rake).

In this case i would override snowie since i assume that people are not playing GTO and will have wrong calling ranges and wrong cbetting ranges.

I would not fold here.

How can u say this and how do you know what GTO is if not via snowie?

Did you really ban me, Rusty? Gratz, I guess.

Thank you for this post.

I think the answer to your question:

a) they don't study a model. They are finding GTO by trial and error. they basically play as many situation as possible and snowie learns which one is more profitable. then they take bots that counter snowies strategies and snowie adjust until it cannot be exploited anymore thus finding something close to GTO.

In this sense GTO = no one can beat it/exploit it. the solution is found by trial and error and approximated by association if no exact solution can be found via the neural network.

b) Well they don't build the strategy, but i guess they could explain how the network works. However the entire strategy is open via the programm which u can download for free for a while.

I find it very useful since you can see the entire range construction. If snowie is not flawed this is the easiest way i can see to learn how to construct ranges.

I don't think the purpose of snowie is to find GTO scientifically but to build the best program for people to learn GTO. In fact they are saying that they cannot get to GTO scientifically but find solutions close to it which this approach.

I cannot imagine how snowie is not effective. they are playing as many possible scenarios as possible and they take counter strategies that do nothing but exploit snowie until snowie has learned how it cannot be exploited.

I understand GTO as the strategy that no other strategy wins against.

So snowie finds exactly that. Again i dont understand why snowie would not find the solution. All the critics always just rambling about how its crap, but i have never actually heard why this approach snowie is taking cannot work.

What really is missing is why snowie construct ranges the way it does. Like why is calling in the QQ example on the flop bad. It however cannot tell us that because it does not know why it does certain things since the strategy is not calculated.

No can developers because all they can do is look at the output of the network.

so what i try to do when looking at spots is try to understand that reason so i can apply it to my real world example. I.e. if we have to check in a spot cuz we can be exploited by getting re-raised but my opponent never does this maybe i can make an exception and bet for value.

No?

ETA: For one thing, you're clearly not banned

Ok, sorry.

ETA: I thought so, but it's okay. Sorry again!

It is very important for the discussion on how close snowie may be on an approximate GTO behavior to examine especially these kind of problems that shock people. Like on post #108 the QQ call being wrong is a massive shocker for all.

We can gain insight however here by forcing snowie to show us how it thinks using that very example.

So those of you that have it try this;

1) First ask snowie by replaying the same hand what it thinks of QQ call of the 3bb open raise preflop. Does it think its bad to not reraise?

2) Same example as #108 only now make hero the first guy and let the 2 callers in as played without giving them hands (keep them as villains).Now at the same exact flop give the villain (now hero) the first player the following hands one by one and see what snowie tells you each time;

AA,KK,QQ,TT,T9s,QJs,99,JJ,AcKc,AcQc,AcJc,KcQc,ATs, JTs, 67s,22 and any hand you can imagine like any combo fdraw+pair or fdraw+str.draw.

Then ask snowie to tell you what it thinks when you cbet the flop with each one of those hands for 50% pot exactly as in the QQ example vs 2 preflop callers.


Keep in a piece of paper note what hands it accepts as ok to cbet with. Make note how bad it sees those above i mentioned that are not good idea to cbet with.

Once we have collected all the hands we can imagine that belong to the opening range preflop of the first player and that snowie accepts as proper cbets at flop, we can go and examine what QQ has as equity vs all these hands as range. We maybe then find that QQ is indeed in trouble as a call if snowie is consistent.

However if snowie allows as cbets almost all of the above hands and then some others if you can imagine any more even say A9s or JKs etc then snowie will be in trouble for lack of consistency.

So allow the possibility that QQ is indeed a funny totally out of the blue bad call here that almost nobody alive would see as such but force snowie to show you why by finding what hands are good cbets vs 2 opponents that didnt reraise you preflop.

Thanks knircky. "Trial and error" can be a method where you indeed don't need to explicitly build a model. However, without any assumption about the underlying function you are implicitly optimizing, you cannot have any guarantee at all that your algorithm will converge anywhere. And if it's just randomly trying strategies and picking the best among them, it's probably the least effective method conceivable. The space of all the potential strategies has a crazy high dimension. Exploring it randomly would require an insanely high number of trials. And even if that was possible, you still would not be able to say anything but 'this is the best I've tried so far'. It's nothing more than trying to play the lottery (and a lottery with a LOT of numbers to be guessed) until you actually win. Except that you can't even really know when you win.

I assume that it's a little more tricky than this. And I'm very inclined to believe that their strategy isn't too bad. But I don't see by which black magic they could come up with any kind of argument making it possible to compare it to GTO.

I really could be missing something here. If so, I'll be very glad if anyone can point it out.

I did some checking for short stack scenarios and ranges proposed by Snowie are far away from push/fold equilibrium. That means solution for the full game is even further...

Example of push chart suggestions? Also see my post above for the QQ case.

Just noticed that thinking of the QQ hand where snowie appaently things in 2NL raising is hugely -EV as well as calling vs on 200nl calling is slightly -EV and both raising and folding 0 EV shows us how crazy the rake is 2NL.

So normally you could raise QQ but at 2NL you have to fold because of the rake.

LOL online poker

actually you can just look up what snowie would have and what it would cbet. You can also look up what snowie puts in villains range.

everything is there for you to see. the entire model is there.


Just because someone here things snowie is stupid does not make it so. i find the fact that no one shows why snowie is bad is pretty decent proof that it is not.

I have been looking for issues with snowie for 2months now and have not seen any (except stupid unfounded 2p2 ramblings).

1. snowie does not calculate pot odds so this is possible

2. snowie is not a good way to determine short stack nash and thats not what its made for

3. snowie is not yet available for tourneys

4. umm could u show us actually what u are claiming. You know i went to the moon last night. was a pretty cool trip. you should try it sometimes.

So if you play GTO there is no strategy that can beat you.

So they claim that snowie plays trillians of hands every day. They have a datacentr with 100 CPU running 24/7.

But they dont ONLY play against rdm opponents they train snowie for a while against all kinds of hands and then have it play against bots that exploit snowie and are optimized to exploit snowie (but might themselve be really bad).

So the same thing the other GTO scientiest use to test their GTOness.

From that snowie becomes unexploitative and no strategy can beat it and it is thus close to GTO.

this way they can show you GTO as well what your ev is against pretty much any situation other players might play. (i.e. playing with 6 limpers preflop)

So snowie would never cbet in this spot itself. it would check 100% of its range.

The only thing it would cbet as preflop raiser would be 22 (bottom set), but 22 are not in its range (it would fold preflop).

3rd caller has a range of suited connectors 45-67s, AQs, JJ,QQ thats it.

So again pretty consistent. Guys yelling how stupid snowie is just don't understand GTO and how poker works.

If it cannot determine correct ranges for 4bb 3-handed poker how can I believe it will provide for 100bb 9-handed? What is your point with tourneys if I look at chip ev mode?
Maybe I look wrong but 3rd caller has no hand he should call preflop...Where is your magic range comes from?

I don't think anyone has said it is bad, just that it is not GTO.