Would it be a valid concept to suggest that a person who can beat an objective standard i.e a bot better than another player might be a superior player...at a technical level, at least.

Im surprised nobody has mentioned this point. Since its trying to play GTO it will have a set of algorithms it religiously follows, a human aware of these algorithms would easily beat it, thereby making it pointless in those situations.
Rusty has made some really interesting points. And is right imo, poker is incredibly complex and will likely never have a GTO solution. God help us if it does.

technically GTO is unexploitable so if snowie was really playing GTO you can't exploit it with knowledge of the algorithms. but obviously snowie isn't GTO

Interesting observation. Thank you!
Basically, computers are fanatic maniacs and should be treated as potential terrorists!

Snowie was demo'd to me by one of the founders. In one hand example, after Hero opened KQo on the button and was 3 bet by the BB to 9x 100bbs effective the Snowie recommendation was to fold, with 4 betting the next best option followed by calling as worst option.

In today's games I would against all but the biggest nits folding seems to be leaving money on the table. This program does not take into account if the BBC is a 90/90/90 or a 2/2/1 it will recommend the same play. If I 3bet 100% it would not adjust.

The founder mentioned that a lot of Zoom players from Montenegro were using the program with some success up to 500nl.

I'm pretty skeptical and don't think it's good for the games as programs like this put off recreational players and I think there must not take into account post-flop playability but use raw equity instead to determine defending ranges to not be exploitable ie probably recommending to defend with hands like A7o instead of KQo in the above example.

Yeah, obv this would be a property of a "GTO" bot.

Knew it.

I think the Snowie guys specifically write on their site that Snowie does NOT take any equities into consideration for his decision-making process.
Whether it's good or bad. Meh, are HUDs good or bad? Was PokerStove good or bad? etc.

I agree that the pokersnowie marketing is, erm, fanciful at best. I've done a fair bit of AI programming, although my focus has been huhu limit hold'em. Even that game is far from 'solved', despite being the "easiest" real game in many respects. (I suspect the best human player will never again beat the best computer at huhu lhe, but that's another matter).

NLHE is at least an order of magnitude more difficult than lhe due to the variations in bet sizes and effective stack sizes. There are major research questions that remain to be addressed. That's before we even talk about multi-way play which is astronomically more difficult as has already been stated.

As far as their marketing goes, they can put up or shut up; there is a yearly computer poker competition for this.

Can I believe they have a strong hu nlhe agent? Yes. Is it better than most tough hu players? Possibly. Is their multi-player agent world-class? I highly doubt it. Can we learn from such strong AIs? I think so, but this is a hard question right now. My gut tells me that the people who would benefit the most are absolute beginners and very advanced players. The problem is a human being may not be able to imitate the way a computer approximates GTO play. For an absolute beginner this probably doesn't matter much, as there are so many huge leaks to plug. Then again, they would probably benefit just as much from playing micro stakes with a $10 'roll. An advanced player is going to be smart about making conclusions from how the ranges work and know which "mistakes" according to the AI are true mistakes and which ones are simply an artifact of how the program works. But it remains to be seen how useful it is for the average players out there.

Yeah, and that doesnt mean its not a scam when they say it is GTO

1) A stripped NLHE GTO poker bot will have at least 1TB (=1000GB) in disk space of different rules applicable to different situation.
2) Good poker players are measured by their ability to exploit the oponent in many different human ways (reads, tells, etc...). And absolutely not by how much their games is GTO approximation.
3) Anyway to beat most of non-expert players today you don't need more than 10 numbers(VPIP, PFR etc...)

It's not really that simple. For one thing, the algorithm only determines the process the model uses to create the optimal formula, and the exact formula would be impossible to reverse engineer unless you had the same model and data set. More importantly though, a truly game theory optimal strategy is not exploitable so the best you could do would be to break even.


I don't know much about game theory beyond the basics, but wouldn't a truly optimal strategy have to be solved by pure math, not using machine learning/simulation? With machine learning or simulations you can only converge to an optimal strategy, you can't actually reach it. Also, current machine learning is really poorly equipped to handle some of the problems that solving the game of poker poses, so I really don't see much progress being made using machine learning without the invention of dramatically different techniques.


As someone who recently got back into poker, I have to say that the recent obsession by poker players with GTO is a bit amusing. As a rule, exploiting the other player's strategy is always going to be more profitable assuming there's a large enough skill difference that it can be done. I'm not saying there's not a place for GTO-style thinking and play in a poker player's arsenal, but its usefulness seems to be overestimated by a lot of players. I find the math/AI part of it an interesting intellectual exercise, but most players would probably be better off, for instance, using empirical evidence from their database to design strategies to exploit other players.

From what I understand about GT yes it has to be purely mathematical (at least that was the case when I studied it at University for a short while). I didn't realise GTO was intended to be "unexploitable", if that's the case then surely it can never be achieved in poker? Most situations are exploitable in poker if you can make a good enough read.

Totally agree with your last point also.

Re 1, good bots don't need to be big.

The bot which has won the Heads-up Limit bankroll event of the Annual Computer Poker Competition many times (a neural net trained to play like the previous year's strongest entrant) is only about 5MB, even though the bots it tries to imitate are usually hundreds of GB in size.

+1. I'd like to see it pitted against (in order) hyperborean, slumbot and tartanian5. We know these bots are attempting to play gto in try their abstractions.

Sent from my GT-I9100 using 2+2 Forums

I very much agree with this. And the method/algorithms behind it would need to be made public (if they are, I haven't seen them referenced anywhere). The University of Alberta poker research group does it right (the academic peer review way), whereas PokerSnowie could very well be snake oil for all I know. For starters, it would be interesting to measure how far from the equilibrium strategy PokerSnowie is. An algorithm for measuring that is discussed here (don't know how it would perform for NLHE, but the point is, it can be measured):

http://forumserver.twoplustwo.com/sh...5&postcount=62

So they need to give us the method, or all the hype in the world won't make me trust or pay for the software. Scientific claims need to be falsifiable, that's Science 101.

you guys seem to forget that it is a commercial program, not a scientific exercise. off cause it is not a bad thing to needle them on this, but my bet is that we won't get any info until some open source project follows in their footsteps. any takers?

Their website uses the phrases "perfect gameplay" and "GTO" a lot. These are phrases with very specific (and scientific) meanings. They say these things to make us pay for a piece of software that may or may not do what it claims. They should prove it to us, or use different words.

ssst ..they may get the idea from this to add one or the other (c) or (r) tag to their claims, complicating matters even more.

The entire strategy is public and can be looked into. You can test snowies move in any situation.

Feel free to prove that it is not GTO. Given the limitation they provided.

There are a bunch of folks here talking about GTO that don't even understand what it is.

again its all out there.

snowie is not a bot and not trying to be good. They have built a way to find the GTO solution of poker via a neuronal network with certain limitations.

the business model is to provide this as a training tool for players at a cost of course.

The bet size is not optimized. So it can only be claimed as GTO approximate.

Sent from my Nexus 5 using 2+2 Forums

Links to the public strategy? If you say "you can enter in hands and see what it recommends" then imo you fail.

Yeah this is totally how validating claims works. Anyone who wants to says what ever they want and if you don't agree you have to prove it. This is why scientists are working around the clock to disprove cold fusion, lizard people, chemtrails, time travel, alien visitations and perpetual motion machines.

Obviously.

have fun scraping strategy from its synaptic mind. this may take a while though. i'd rather send them a mail and ask for a brain dump ;-)

i wonder if the following two points find agreement here. i don't mind it being around if:

1 - it is made sure that it can not be used in-game (not some semi working solution, but something that is bullet prove)
2 - it can be used to rank players by some (provably) objective skill measure so we can in the end detect it (or it's mindset) and ban it from tables we don't want to see it at

Do you have a link?

That's not how the scientific approach works, it's the other way around. They label their strategy "GTO" without any (afaik) hard evidence that it's even close (given limitations, which lead to a simplified game, which has some GTO strategy that they are trying to approach). Since there are algorithms for measuring that (see link in my previous post), they have a bit of work to do before they can market the program so boldly, imo.

This is true. For starters, GTO isn't even well defined for multiway pots (I have limited understanding of this, but it has to do with alliances that can be formed between players, right?), so it's a little dubious to claim "perfect GTO play" in marketing for that reason alone.

collusion is ruled out.

i am too lazy to dig it out, but this was the gist of the reply to some weakish prove attempt on why GTO can not exist in multi player situations and thus sno**e could always be beat in theory (discussed in this subforum not long ago)