Not sure what you consider 'same general structure', but in backgammon Snowie got very strong using their neural net learning.

At the end of many backgammon games there comes a time when there are no more opposing checkers between you and bearing off. In these "no contact" positions there is an exact solution for bearing off checkers. Snowie was never able to perfect it, while Hugh Sconyers, a noted player provided a set of CDs with the database solution to these spots.

So, would a "no contact" bearoff be 'of a same general structure', much like "short stak strategy" might be a smaller subset of NLHE ??


which

Was Backgammon Snowie generally considered "weak" in these bear-off positions? Or still very strong, just not quite perfect?

If it was still strong, that sounds to me like a similar situation that we face here in poker: Backgammon Snowie likely doesn't play perfect in either case, bear-off or full, but we are only able to pinpoint mistakes in the simple subgame where we can solve the game using other methods.

Still strong, but provably wrong.

Much the same as the math based Push Bot strategy may never be duplicated by PokerSnowie (maybe just a close approximation)

which

I've registered for the 10 days free trial of PokerCoach. The hand importer can't handle Boss Media format (used by EspaceJeux in Quebec). That's a bummer.

I played a couple hundred hands HU against Pokersnowie and am puzzled by something. I enable live evaluation to get immediate feedback on my decisions. I choose a line and compare it to Snowie's suggestion but I then always stick to my initial choice in order to get a proper assessment of my play at the end of the session. What puzzles me is the 'optimal' bet sizing recommendation. Oftentimes, I have chosen a different bet size so I check the difference in EV between my choice and Snowie's recommendation. It often is the case that my initial choice has a higher EV by a significant amount (close to 1bb). Why is Snowie recommending some specific bet size when other bet sizes have a higher EV according to him/it/her ??

As far as i understand, Snowie is not capable of splitting its range into multiple sizings and instead uses the single sizing that results in the best overall EV for the range.

Ah I understand! If Snowie would always choose the highest EV bet size, then its lines would become unbalanced and exploitable. So, that's one significant difference from true GTO equilibrium where any single strategic decision must maximize EV against another GTO opponent.

That means I am free to choose the bet size that has the highest EV, according to Snowie -- while playing against him -- while being aware that doing so might throw my lines out of balance and make me exploitable by decent hand readers.

Thank you very much!

Snowie will only chose only one size for it's whole range in any given spot. But this is done because it's considered to complex for the program to try to balance multiple ranges in the same spot. It is possible in game theory to have multiple bet sizing in the same spot with each range being balanced.

Indeed, that's what I understood. I also pointed out that when a strategy is optimal (GTO) then a bet size never would be chosen for the sake a balancing the range in the spot where this bet size is used if this would involve sacrificing some EV against a GTO opponent. That would mean that the strategy isn't maximally exploitative against its nemesis and this would contradict the condition of equilibrium.

In practice this means that when there is a bet size that has a higher EV than the one recommended by Snowie, we ought to analyse what accounts for the lower EV of the recommended size and, maybe, consider some mixed strategy that would increase our EV in that spot without compromising the balance of our lines too much.

I've played a few hands vs various AI HU bots and enjoyed playing snowie and found it interesting.
I think the whole thing is a good idea in general but I think snowie is spewie in certian spots.


I played 1st session at $1/$2 HU 100BB AutoReload:

Me: VPIP: 76.82% PFR: 73.91% 686 hands
Snowie: VPIP 68.22% PFR 55.40%

Blunders 35(I checked these and I agree with a few but am really happy with stuff I did which snowie didn't like, eg:snowie making VB on river I mix my Value Hands and Bluffs with Raise/shoving some spots. Think it caught me once where after I thought I had made a spew in that particular spot thinking about it)

Profit: $1.54/Hand Total Profit $1,059

Snowie rated my play Intermediate Level with 18.53 Error rate.

I thought I played well(tired)but definitely made 2 or 3 bad mistakes.

Session 2 I played $1/$2 HU 100BB AutoReload:

Me: VPIP 78% PFR 76% 100 hands
Snowie: VPIP 71.73% PFR 51.52%

Blunders: 3(Checked all spots and really think snowie is fos, think some of it's lines where super exploitable against my lines, caught him 2x potting river on missed draw, and some other smaller mistakes) I think I didn’t make any mistakes apart from 3B folding 66 v a 4B. I ranged snowie after with all hands I saw him 4B and think you can profitably jam 66+ATo+A8s+ and maybe KQ too in the same spot. In total I have seen snowie 4B fold v a 5B Jam 4 times over 786 Hands.)

Profit: $2.27/Hand Total Profit $297

Snowie rated my play Advanced Level with 13.36 error rate

3rd session v Snowie $1/$2 HU Freezeout 100BB:

Me: VPIP 70% PFR 70% 20 hands
Snowie: VPIP 65% PFR 38.65%

Profit: $10/Hand Total Profit $200

Blunders: 0

Snowie rated my play Advanced Level with 13.52 Error rate. (Not sure how many errors I can make during 20 hands which snowie got caught bluffing when I had the mortal nuts. GG snowie)



My conclusion. I think the program is called snowie because when I play it it feels like christmas!

PartyScout

I only skimmed this thread and have not yet downloaded Pokersnowie myself. I saw several comments about how they have not or would not enter it in the annual computer poker competition. My question is this: Is there anything stopping someone from entering Pokersnowie for them? That is, enter a program that simply queries Snowie on every situation that comes up. Or is it not designed to allow this?

In the ACPC, you have to upload a program that can run independently; you aren't allowed to connect to the internet:

Competitors are reminded that this is a competition about artificial intelligence in poker, and any attempt to manipulate the outcome by abusing TCP/IP (by trying to packet sniff, communicate with other bots, connect to the internet, or use denial of service attacks) may result in disqualification.

Bad preflop I know, just been messing around against snowie trying to see when I can make him fold. I played a similair spot IP after this one and got him to fold AA when I jammed 5s2s on JcQcQx9c ,I float flop after snowie 5b me OOP PF, snowie half pots flop then c/f turn when i jam turn. 200BB deep start of the hand

Anyway, snowie thinks after we see this flop we should fold 100% hahahhaha. snowie is a nitfish! Snowie would get destroyed by every single bot at the annual computer competition. I like this program now, but I really think it will only make me play bad poker(as I have deviated far from my usual game to exploit snowie) Fun though!

Snowie recommends that you should fold K9o (with a K high flush draw) on this particular spot against Snowie's own range. That's because you EV in that spot is -34bb against Snowie's range. Why would you call any more than 0% of the time? For the call to be part of a mixed strategy, its EV would have to be zero.

I am sure there are HUNL bots that are stronger than Snowie, but it isn't obious that this is an exploitable spot. Snowie will not necessarily get in that spot in the BB with the same range as you do. Even if he gets there with this particular hand, he may still perform OK with his whole range while folding this particular hand to a shove against any opponent.

Does HM2 work on PS? If not, how would you be able to see PS' HUD stats?

Hi Poker Snowie. Can you please explain (confirm it's not an error) your optimal 3B squeeze ranges 50BB deep, for FRNL 9 handed, from the BB, after UTG opens and the HJ calls? At the moment Snowie says that I should 3B 72s 21% of the time here and 32s 35%?

At 25/50, 50BB stacks, nlh 9 handed, UTG goes $112.5, folds to the CO who calls the $112.5 and then it folds to our BB and we are supposed to 3 bet 72s 21% of the time? There are other weird hands in its range that seem to weak to raise as well. Can you please elaborate?

I've just downloaded the trial version of PS and tried the HU challenge mode.

Is there only one difficulty setting on the trial version? Are there different difficulty settings in the full version?

Overall, I found the challenge mode a bit easy, PS play straight forward and didn't put me into too many tough situations barring the odd 3bet and flop check-raise.

Maybe it's because I've only played a few hundred hands so far but my overall first impression is that I don't find myself put into tough spots by PS as I do compared to when I play vs regs and even aggressive fish.

Not doubting that I won't learn to improve using PS, because I do think it can help my game, but just not entirely sure whether it would improve me as much as I hoped as I don't feel that PS truly reflects the range of opponents that I play against.

is it actually possible to see snowie's ranges for each strategic option for an arbitrary spot? so assuming snowie plays vs itself (using its betsizing restrictions) in a 100bb hu spot with an arbitrary board, say KdTc2s4h7d, and arbitrary lines taken by snowie sb and snowie bb that are within the besizing space, say sb raises pot (3bb), bb raises pot (9bb), sb calls, flop bb bets half pot (9bb), sb calls, turn bb bets half pot (18bb), sb calls, river - can we see what the river starting ranges for sb and bb are and how they play them given their options? furthermore, is it possible to see all other ranges at each decision node?

if that's the case it would be very intersting to have a look into those ranges and frequencies, eg the preflop 3betting range. additionally, it would be easy to plug in some river spots and see whether those strategy pairs actually build a stable equilibrium (in a spot that is most likely off-the-equilibrium path even in the truncated game).

Having just read through the entire thread, the pertinent issues seem to me to be:

1. Does PokerSnowie understand that it is not the case that a neural net is not guaranteed to (as they say) "automatically lead to a GTO approximation," that this is a claim that requires mathematical proof? I don't think anyone doubts that their algorithm leads the program to play better, but to claim that it converges to the GTO strategy is merely an assumption.

2. Does PokerSnowie understand the precise mathematical definition of a Nash equilibrium and, consequently, what is required in order to claim that a strategy is an approximation of the GTO strategy?

3. Assuming PokerSnowie actually is confident in their claims about the strength of their program, is there any legitimate reason that is keeping them from measuring the exploitability of their program (which, as FullyCompletely clarified, is feasible in the sense of getting a lower bound) or entering the poker bot competition? It was said that the latter is being taken into consideration. But it's not as if undertaking either of these would be particularly difficult or time consuming. If they really think their program is the world's best, what reason could there be not to do either of these?

I for one would love to hear PokerSnowie's thoughts on these matters.

poker snowie stopped answering questions.

To me, seems all a marketing thing...
I don't see any value selling it to the pubblic, for a few bucks, instead keeping it secret and make millions at the table for its owner

Snowie isn't close to GTO and is quite unbalanced....

i wasn't implying that, i was asking a question..

In a line that is off-equilibrium at least one of the ranges needs to be empty,...

I'm pretty sure you can see Snowie's ranges in every spot.

thanks!

? you either didn't follow me, or i can't follow you