I wasn't writing about my personal preferences. I was simply pointing out that the sense in which snowie's strategy is "publicly available" is limited and, as a result, certain tasks that are central to the discussion in this thread, such as measuring the strategy's exploitability, cannot be performed.

he already owned you right there, and yet you don't seem to get the (very obvious and important) argument he and many many others throughout this thread are making.

TrashTalk ftw

I think there are two thing at work here:

1. Snowie is made for humans and not computers. So yea if u wanna use it against computers it's not been designed that way.

2. The tree snowie has must be much much more complex than bots, I'm sure bots solve the problem by design, while snowie solutions is derived like ants, chance, evolution etc.

Think about making a chess board by design (no tile may have same color). Takes like 64 steps. By evolution it would be something in the 5k range and by rdm chance I think it takes trillions of tries. I think snowie is the 2nd example.

Well have u looked at its strategy and range overviews. It sounds like you think the only way to look at their strategy is via scenarios?

It's more like snowie shows you anywhere u can go from where u are and where you could have come from.

But you are right it only shows one step. Anything more would not be readable for humans though clearly for computers.

1. I am aware that the snowie program that can be downloaded is not designed to play against bots. I wasn't suggesting that they do that. I'm suggesting they make snowie's strategy (which they already have stored in a file somewhere) publicly available so that other people can see how that strategy plays against other (bots') strategies. Note that this wouldn't be difficult once the snowie's strategy can be obtained.

2. When you say "must be", I interpret that to mean that that is your best guess. In other words, you haven't checked this with either poker snowie or the owners of other bots, and thus you don't know this as a fact. That's not a criticism; just wanted to make that clarification because you stated it as if it is obvious (and I'm sure many here would disagree).

The (file) size of a strategy is not necessarily dependent on how the strategy was derived. I think two significant factors in determining the (file) size of a strategy are the number of available bet sizes (and I think most bots have more than 3, the number snowie has) and the way the strategy groups situations (I don't know much about this for either snowie or other bots). Do you have any reason for thinking that a strategy derived through neural nets is larger than one derived through algorithms such as CFR?

You used terms like design, chance, and evolution to support your argument, but afaik those aren't well-defined terms in computer science. To increase the validity of your case, you might want to use more common and precise terminology and perhaps link to some publications that support your claims.

I agree it would be beneficial if they were to make their strategy available for machines. However this is not part of their mission. So I think its not quite fair to basically tell them to do something that is not really their business.

Snowie could provide an API, but again this is not their business. They are not a science shop they provide a training tool for poker players, humans. The other bots have a very different objective (i.e. u cant use them as humans to learn anything).

That said if you know a strategy is GTO you could very easy compare if snowie follows it since its totally open. It would almost not be fair to have snowie compete with bots since the bots know exactly what snowie does in any point while snowie does not. Of course we can argue snowie should not care if it really were GTO :-)

You are correct. I have no clue. I follow snowie out of interest. I think it is revolutionary new technology that will shape how we think about poker. i dont any snowie people i have just followed them.

I dont think snowie has a strategy file per se. I think the snowie strategy is likely very simple: just do whats most EV, but it looks up every path from the decision tree. I would assume a bot is not going to write the decision tree but instead calculate what to do intelligently. But here i also have no clue i am just guessing. But snowie takes a different approach to GTO. So since snowie has no intelligence per se and knowledge of the game and does not do any pot odds etc calculations it just follows a path that yields highest EV. Just like an ant. As such it must be able to look up the entire tree I would think. We know snowie has a data center with 100 CPUs and im sure their data store is HUGE and certainly not capable to be stored in a file. There are some on this forum that have posted the number of decisions and file sizes etc that would be needed and they were always out of this world. Snowie has simplified this but it still must be insane. Also the fact that they strategy changes over time shows this.

The approach that snowie is taking is an evolutionary approach vs one by design. This term comes from the battle in science between the folks that believe in evolution vs creation by design (GOD in this case).

So evolution happens rdmly but follows certain rules. I.e. if we were to fill a chess board by design (i.e. a human) you could tell the person to make sure every tile is different. the designer would be able to color the board in 64 steps because it would not make any mistakes.

Evolution would color the first board and then rdmly the next. And then it would trow away the results that don't work and keep on doing this producing lots of mistakes and having to to redo many steps. Eventually it would come up with the correct solution after about 5-6k tries and semi-randomly (it keeps parts of the solution that works).

a total rdm try (i.e, coloring every tile in board rdmly and then checking if the board is valid would take trillians of tries or so.) So evolution can be stupid and still relatively effective (i.e. create species like humans that are pretty intelligent rdmly)

Poker is so complicated that you cannot solve it by a designer (or at least we have not been able to) so now snowie takes this evolution type of approach. Its dumb very expensive but also very effective because you dont need a designer (very smart developer).

Ants work the same way. One ant finds a good path and more and more follow. the more that follow the better the path. So over time the best path gets found this way. So pretty dumb animals can solve big problems this way. Your brain learns this way also.

hence i dont think the data can be published via file. And this is why there is no GTO strategy file on your PC when u download snowie. But the snowie connects to the server and looks up the strategy there.

If snowie were to make its strategy public it likely would only work via an API but best way is prolly to just play against it and test it that way.

BUT again i have no clue what i am talking about here (just guessing)

Since part of their mission is marketing, I think it is very much their business to demonstrate their claims such as being the best bot in the world by actually playing against other bots.

Not sure what you're trying to say in the last paragraph. When bots play each other, they don't look at the strategy files of other bots (even if they were available); that would defeat the purpose of the competition. Of course, snowie is the same; it doesn't look at the "strategy files" of human players (although these don't exist in the first place). Snowie would not be at any kind of disadvantage for disclosing its strategy in order to pit it against other bots. The only way it would disadvantageous would be if someone took snowie's strategy and computed a maximally exploitive strategy against it, but that's uninteresting and not what's being suggested here.

Snowie only traverses a subtree (only 3 bet sizes, and whatever other groupings it uses), so this doesn't guarantee to yield the highest EV – it only guarantees to yield the highest EV in the subgame.

The file sizes are indeed large, by they are not unmanageable, as demonstrated by the poker bot competitions where bots with strategies of "insane" sizes were pitted against each other.

A neural net is also an algorithm and is thus made by design. Different neural nets can be programmed, and they will converge to and yield different strategies. Who knows – there may even be neural nets that guarantee to converge to GTO strategies. Currently, however, there are no such neural nets for poker afaik. That being the case, it is also true that snowie's "evolution type" also cannot solve NLHE. It's strategy is continually getting stronger, but it is not guaranteed to converge to a Nash equilibrium strategy (i.e. solve the game).

What you wrote prior to this sentence seems unrelated to whether the file can be published or not. I'll just note that if we're talking about computational possibility, it is possible to publish any file of any size.

does anyone put emphasis on the rating of your play that snowie does, like between 1-30 or whatever it is? I've heard 6 is a good mark for a game that includes a good mix of GTO and exploitative strategies.

Lol this post is so absurd it's not even funny.

First you claim Snowie has helped you understand GTO. THEN you claim you use this knowledge to beat 25NL.

Do you even know what GTO is? Even if you can play a Nash strategy, it's only optimal IF all other players also play optimally. At 10/25NL, everyone is so far from optimal it's not even funny. Using Snowie as a way to beat micro stakes is just completely absurd and GTO approaches are NEVER optimal again opponents who basically suck ass at Poker (AKA pretty much everyone at micros).

On another note, claiming Snowie is a good program because it helped you improve at micros is also ridiculous. It's a GTO approximation. How close that approximation is is the real question. Pretty much any half-assed approximation can show you enough about how to not play like a ****** to let you beat micros. Snowie can be miles and miles away from GTO and still perform well at micros just because everyone is so bad there that any somewhat fundamentally sound approach is likely to win.

e: FWIW, here are my general thoughts on Snowie. I did an undergrad thesis on Poker agents, and have alot of experience in Comp Sci / Game Theory. Basically, Snowie is just a neural network which played trillions of hands against itself and learned what moves are +EV and -EV. For the obvious and frequent spots like raising AA pre, this approach performs well since it doesn't take many trials to learn that raising the nuts pre is +EV. However, looking at papers such as http://poker.cs.ualberta.ca/publicat...rt-nl-size.pdf we can see that there are SO many different situations in heads up poker alone that to try to brute-force a complete GTO solution would require way more computing power than anyone has. Modern computer systems are nowhere near powerful enough to do this in any reasonable time frame. There are so many different game states, you would need a computer the size of Mars along with some innovative technology that helps you overcome common computing problems like the power wall, memory wall, decline of Moore's law, etc. in order to come close to solving HUNLHE.

To put it in perspective, some of the best people in the poker AI field from Alberta have come close to solving GTO for HU limit hold'em, and even then they don't have a perfect solution yet. NLHE game trees are exponentially larger than LHE counterparts due to all the different bet sizing options. Even with abstraction, this is way too much to brute-force through.

Not only is Snowie claiming it somehow brute force solved NLHU which nobody has even come close to doing, it's also claiming it solved 6-max which is completely absurd. Yes, it may have gotten an approximation but that approximation is probably so so far from GTO that it doesn't even come close to playing optimal poker.

What does a solution look like? I am familiar with matrix GT games... but I couldn't imagine the solution for even one single poker spot

You should ask for a tuition refund IMO. Your post is all kinds of wrong.

I will expand if you reply that you will read my response and reply to it.

I didn't proofread or word it very well so there are likely some errors. I'm interested in what you have to say though

Snowie is at it again...

"PokerSnowie has learnt GTO using..."

http://www.pokersnowie.com/blog/2014...y#.U4kNZi_j6DZ

Yeah I pretty much thought most of Any2's post you said had a lot wrong with it was correct. What was incorrect?

How can you do a thesis at university if you can't even read...
It's easy to bash on pokersnowie but did you try it allready? If you are that kind of wizzard that you say you are, why are you still playing 10NL??

Funny guys on forums you must love them!

I don't have to try it. I know it's computationally infeasible to come even close to finding a GTO solution for deep stacked heads up NLHE. The fact that Snowie is claiming to play GTO in 6-max situations is just ridiculous, I have better things to do than test a program that I already know is BS.

Also, what do the stakes I have play have to do with Snowie? I play poker casually on the side. I make a bit of money donking around at micros but maybe play like 5k hands a month on average. Most of my attention is focused on my studies.

I just saw that it advocates checking 99 behind in position on 985tt board in a heads up pot. The reasoning they give is that "it's a bad board texture". I would think that since we have big suited hands in our UTG range we would want to balance our semi-bluffs by also betting top set here. With 99, we would want to get value from lower sets and 98 for two-pair before a wet turn card comes and kills our action. We would also want to get value from any draws villain may have. Can anyone explain why checking behind here might be optimal, or is Snowie just FOS?

snowie does some "weird" stuff and I've questioned a lot of its plays myself. with that said, I do think its a good training tool and I could teach an alien to play well pretty quickly w/ poker snowie. yess I know snowie is not GTO and we don't even know what "close to GTO" should be - however, that's just sales bs and we get that and yes they are wrong to say they are gto but that's not up for debate.

I've been playing 200-1knl (probably average 600nl) for years now and also studied game theory/machine learning while getting my undergraduate computer science bs. I'm not a wizard but I'll say I agree with pretty much everything he said in that post.

Talking about people bashing snowie while not trying it is again not understanding burden of proof. It's like the choice center defenders in nvg saying if you don't try the course you clearly can't make a claim that it doesn't change lives. The burden of proof is on the ones making the claims. Snowie hasn't proven anything so I'm going to bash it for making claims that are false and making claims they can't prove. The way you shut people like me up who are bashing snowie is you prove the claims that snowie make that give it value. Either it's gto or it's strategy is sound enough to beat x limit, or whatever. The way you show people you're an idiot who doesn't understand logic is to make ad hominem attacks and show that you don't understand burden of proof.

I'm right and everybody else is an idiot...
You must be a smart man!

Thanks zach, glad to hear I at least got some things right and my tuition wasn't completely wasted.

If we ignore for a second the fact that solving GTO for NLHE is pretty much impossible and pretend that Snowie does in fact play "perfect" poker, then the whole thing still seems really strange. If someone out there magically discovered a perfect strategy for NLHE heads up and 6-max, why would they ever release it publicly for a measly ~$200/year per customer? That knowledge is worth millions of dollars, but only retains its value as long as it's kept secret. If you have a program that can tell you the GTO play in every cash scenario, just use that when playing on Stars or any other major room. Programming an interface that captures hand data from the screen and inputs it into your GTO program which outputs the optimal play is trivial for any somewhat competent programmer. You could then use this to crush absolutely every stake online. How sick would it be to know you have a strategy that cannot lose long-term? You could build a roll extremely fast and then sit in pretty much any high stakes game and beat it. Even winning just 2bb/100 playing 2k hands a day of 5k NL you make $2000/day or roughly $500k/year, and that's pre-rakeback and assuming you only play 5 days per week. If you then move up to the nosebleeds with this strategy and beat the best regs at even 1bb/100 you make an absolute fortune. Expand this further and program a team of bots that plays this strategy at mid-high stakes across multiple networks and you're actually printing millions per year. After enough experience and study of what your program tells you to do, you'll start to get a feel of what GTO poker entails, and if you get good enough you could memorize the general gist of the strategy and employ it in live games and crush there too.

So, you have to ask why on earth the people behind Snowie would ever reveal anything about the bots strategy if it really was GTO. Once you reveal the strategy to the world, anyone can start using it. Once everyone starts using it, everyone breaks even pre-rake since poker without rake is zero-sum. Then, after rake, everyone loses and you've basically made your program completely worthless. It's completely illogical to sell a GTO poker strategy to the world for $200/year per customer. The only explanation why someone would do this is if the program wasn't actually that good and they just wanted to take advantage of a few naive donks.

I estimate you're age at 11...

Oh, another ad hominem attack with no concrete counter-argument. How much are Snowie paying you to do this? Frankly it's really a very poor effort.

Your. You'll learn that in 6th grade.

Snowie is awful.... I'd pay money to play snowie at 5-10 for the rest of my life.

Yes it's good for players to improve to beat like 50nl at best.

If anyone thinks it'll help them beat decent stakes you're fooling yourselves.

If snowie wants to challenge me for real money i'd love to, and i'm only a mid stakes player =].