Pokersnowie question
10-01-2016
, 11:39 PM
Quote:
It seems like Snowie offers advice that is pretty bat**** crazy in some spots. After reading one of Arty's posts in BQ, I wanted to see how Snowie deals with having a narrow, capped pre-flop range on bad boards. 100NL 6-max UTG opens to 3xBB and faces a BU 3-bet to 10.5xBB. UTG calls with a range of 99-JJ, AJs-AQs. It seems to me that this range is going to be very hard to play on a large number of boards. On Axx boards our strongest hand is AQ when BU can have AK, AA and some light 3-bets which hit 2P/sets. On Kxx our strongest hand is second pair. On any board 8-high or lower our opponent has QQ+ and we have no sets. We can get a board which smashes our range like JT8 and effectively win nearly the whole pot, but these are rare.
So I looked at the A
7
3
and gave UTG A
Q. UTG checks AQs along with it's whole range here, which seems clearly correct. BU bets 1/2 pot 36% of the time, which seems low to me and I see that Snowie is only betting 2P+ and bluffs on this flop. When UTG bets 1/2 pot Snowie defends AQ/AJ only and folds 71% of the time. Now things get weird. Turn is T and snowie leads 1/4 pot for an EV of 23.45 
This seems blatantly incorrect as the turn adds no strong hands to our range, our range is face-up so our opponent's range is polarized. The turn seems like an obvious check. Of course BU only folds to this bet 6% of the time, but his defending strategy is ridiculous. Snowie says raise 1/2 pot 30% of the time with a range of only bluffs, and call 63% with a range of 2P+ and some floats. When the opponent raises Snowie says to never fold, which makes sense if the opponent's range is 100% bluffs, but in practice and theory this line is nuts.
After BU calls and the river is the 4, Snowie says the EV of betting 1/4 pot again is equal to the EV of checking at 30.38, which is obviously wrong because the opponent can just jam on us with a balanced range. Snowie then says BU should raise 1/4 pot with all hands that can beat AQ, but no bluffs. Obviously the correct bet size here is all-in and we should have a balanced number of bluffs. Snowie still thinks we should call the raise sometimes even though the opponent's range has no bluffs.
Going back, if the flop goes x/x and the turn is the same T, Snowie says check TT and bet the rest at 1/4 pot, which is again nuts as it makes it impossible for us to get value with TT and our betting range is weak.
IDK wtf is going on. Does Snowie just completely suck at playing narrow ranges?
So I looked at the A
73 and gave UTG AQ. UTG checks AQs along with it's whole range here, which seems clearly correct. BU bets 1/2 pot 36% of the time, which seems low to me and I see that Snowie is only betting 2P+ and bluffs on this flop. When UTG bets 1/2 pot Snowie defends AQ/AJ only and folds 71% of the time. Now things get weird. Turn is T and snowie leads 1/4 pot for an EV of 23.45 This seems blatantly incorrect as the turn adds no strong hands to our range, our range is face-up so our opponent's range is polarized. The turn seems like an obvious check. Of course BU only folds to this bet 6% of the time, but his defending strategy is ridiculous. Snowie says raise 1/2 pot 30% of the time with a range of only bluffs, and call 63% with a range of 2P+ and some floats. When the opponent raises Snowie says to never fold, which makes sense if the opponent's range is 100% bluffs, but in practice and theory this line is nuts.After BU calls and the river is the 4
, Snowie says the EV of betting 1/4 pot again is equal to the EV of checking at 30.38, which is obviously wrong because the opponent can just jam on us with a balanced range. Snowie then says BU should raise 1/4 pot with all hands that can beat AQ, but no bluffs. Obviously the correct bet size here is all-in and we should have a balanced number of bluffs. Snowie still thinks we should call the raise sometimes even though the opponent's range has no bluffs.Going back, if the flop goes x/x and the turn is the same T
, Snowie says check TT and bet the rest at 1/4 pot, which is again nuts as it makes it impossible for us to get value with TT and our betting range is weak.IDK wtf is going on. Does Snowie just completely suck at playing narrow ranges?
10-02-2016
, 09:32 PM
Snowie says call {ATs-AQs, 77-JJ}.
Flop A
73, now Snowie wants to bet 1/4 pot with 88-TT a small percentage of the time.When BU is checked to Snowie bets more often now. 45% of the time. It uses a mixed strategy with AK instead of always checking it now, except AK with BDFD is still checked always.
When UTG faces a bet of 1/2 pot it now defends AQs-ATs, 77 by calling, and raises 77 50% of the time. No bluffs to go with 77 in its raising range.
On the T
turn Snowie now checks it's 77/AT sometimes, and bets the rest of it's range at 1/4 pot. When the BU faces this bet it still recommends raising with a range of only bluffs, 39% of the time now.When BU calls and the river is the 4
, Snowie now recommends a mixed strategy with all hands except 77 which is always checked. It still leads with the 1/4 pot sizing. Now when Snowie gets shipped on (still all value) it wants to fold 80% of the time (opponent risking 75.25 to win 64 in the scenario I set up)Snowie thinks BU should call twice with hands like KT instead of turning them into bluffs, despite them not beating anything in UTG's betting range.
Still some wacky stuff...
10-03-2016
, 07:34 PM
i.e. When the neural network was training against bots with bad/random strategies, it would find that some hands are +EV calls on the river against those imperfect ranges, so those hands get added to the database as "+EV calls", although they don't make money against the (supposedly) "optimal" range that Snowie itself would bet. I'd guess that Snowie needs to play against itself more. :/
I presume that the problems are compounded when the database is augmented with (rather than replaced by) new results using new bet-sizes, since a database of results (particularly the EV figures) from training with half pot bets isn't going to be accurate once quarter-pot bets (and entirely new strategies for the earlier streets) are also allowed.
Snowie is still learning and still does "weird" stuff. It probably always will. The game is too big for it to have learned the solution for every possible spot and every possible strategy.
One little thing that I'd add is that while there are lots of things that Snowie does that look weird to humans like us, I think it's probably true that GTO poker actually IS quite weird. I wouldn't be particularly surprised if min-betting into a 20bb pot is sometimes optimal. The bots at the annual computer poker AI competition were doing that sort of thing years ago, and we also saw "weird" stuff like Claudico betting 18x pot with T5 high. If you're using mixed strategies that involve very low frequencies with some combos, you can probably do just about anything and call it GTO.
Last edited by ArtyMcFly; 10-03-2016 at 07:44 PM.
10-04-2016
, 01:03 AM
Quote:
Doing the same thing with a 2xBB open and facing a 7.5xBB BU 3!
Snowie says call {ATs-AQs, 77-JJ}.
Flop A73, now Snowie wants to bet 1/4 pot with 88-TT a small percentage of the time.
When BU is checked to Snowie bets more often now. 45% of the time. It uses a mixed strategy with AK instead of always checking it now, except AK with BDFD is still checked always.
When UTG faces a bet of 1/2 pot it now defends AQs-ATs, 77 by calling, and raises 77 50% of the time. No bluffs to go with 77 in its raising range.
On the T turn Snowie now checks it's 77/AT sometimes, and bets the rest of it's range at 1/4 pot. When the BU faces this bet it still recommends raising with a range of only bluffs, 39% of the time now.
When BU calls and the river is the 4, Snowie now recommends a mixed strategy with all hands except 77 which is always checked. It still leads with the 1/4 pot sizing. Now when Snowie gets shipped on (still all value) it wants to fold 80% of the time (opponent risking 75.25 to win 64 in the scenario I set up)
Snowie thinks BU should call twice with hands like KT instead of turning them into bluffs, despite them not beating anything in UTG's betting range.
Still some wacky stuff...
Snowie says call {ATs-AQs, 77-JJ}.
Flop A
73, now Snowie wants to bet 1/4 pot with 88-TT a small percentage of the time.When BU is checked to Snowie bets more often now. 45% of the time. It uses a mixed strategy with AK instead of always checking it now, except AK with BDFD is still checked always.
When UTG faces a bet of 1/2 pot it now defends AQs-ATs, 77 by calling, and raises 77 50% of the time. No bluffs to go with 77 in its raising range.
On the T
turn Snowie now checks it's 77/AT sometimes, and bets the rest of it's range at 1/4 pot. When the BU faces this bet it still recommends raising with a range of only bluffs, 39% of the time now.When BU calls and the river is the 4
, Snowie now recommends a mixed strategy with all hands except 77 which is always checked. It still leads with the 1/4 pot sizing. Now when Snowie gets shipped on (still all value) it wants to fold 80% of the time (opponent risking 75.25 to win 64 in the scenario I set up)Snowie thinks BU should call twice with hands like KT instead of turning them into bluffs, despite them not beating anything in UTG's betting range.
Still some wacky stuff...
As far as hoping that they explain some of the stuff, it's never going to happen as they don't know enough poker to explain anything.
http://www.thehendonmob.com/search/?...search=Olivier
There is nothing to document that Olivier Egger has made money playing poker.
http://www.thehendonmob.com/search/?...hannes&search=
There is nothing to document that Johannes Strassmann has made money playing poker.
Neither appears in thehendonmob.com database.
I would like for them to post their screen names on some poker sites so that we can check with sharkscope.com and see if they actually play poker.
They are trying to cash in on the success of their Backgammon bot but that does not translate well to real world poker.
And to the poster stating that Snowie beat Jungleman that is not true at all. Jungleman beat the hell out of snowie HU. Don't take what some of these cheerleaders say at face value.
I have seen a video of player at the 2016 WSOP say that they played a hand according to Pio solver. If there is a video of pro saying they played a hand according to Snowie I will like to see it.
Last edited by wwwin; 10-04-2016 at 01:32 AM.
10-04-2016
, 02:26 AM
10-04-2016
, 03:50 AM
Quote:
One little thing that I'd add is that while there are lots of things that Snowie does that look weird to humans like us, I think it's probably true that GTO poker actually IS quite weird. I wouldn't be particularly surprised if min-betting into a 20bb pot is sometimes optimal. The bots at the annual computer poker AI competition were doing that sort of thing years ago, and we also saw "weird" stuff like Claudico betting 18x pot with T5 high. If you're using mixed strategies that involve very low frequencies with some combos, you can probably do just about anything and call it GTO.
About the other stuff you said, I guess I didn't realize how Snowie works. It actually sounds like some of the lines it takes are exploitative against the opponents it has trained against and not even designed to be an approximation of GTO. I thought Snowie was supposed to be a pseudo-GTO engine. That doesn't seem to be what it is.
The sample size issue makes sense.
10-11-2016
, 10:57 AM
Rake also affects the ranges Snowie plays.
10-13-2016
, 01:44 PM
10-14-2016
, 03:01 AM
Quote:
Doing the same thing with a 2xBB open and facing a 7.5xBB BU 3!
Snowie says call {ATs-AQs, 77-JJ}.
Flop A73, now Snowie wants to bet 1/4 pot with 88-TT a small percentage of the time.
When BU is checked to Snowie bets more often now. 45% of the time. It uses a mixed strategy with AK instead of always checking it now, except AK with BDFD is still checked always.
When UTG faces a bet of 1/2 pot it now defends AQs-ATs, 77 by calling, and raises 77 50% of the time. No bluffs to go with 77 in its raising range.
On the T turn Snowie now checks it's 77/AT sometimes, and bets the rest of it's range at 1/4 pot. When the BU faces this bet it still recommends raising with a range of only bluffs, 39% of the time now.
When BU calls and the river is the 4, Snowie now recommends a mixed strategy with all hands except 77 which is always checked. It still leads with the 1/4 pot sizing. Now when Snowie gets shipped on (still all value) it wants to fold 80% of the time (opponent risking 75.25 to win 64 in the scenario I set up)
Snowie thinks BU should call twice with hands like KT instead of turning them into bluffs, despite them not beating anything in UTG's betting range.
Still some wacky stuff...
Snowie says call {ATs-AQs, 77-JJ}.
Flop A
73, now Snowie wants to bet 1/4 pot with 88-TT a small percentage of the time.When BU is checked to Snowie bets more often now. 45% of the time. It uses a mixed strategy with AK instead of always checking it now, except AK with BDFD is still checked always.
When UTG faces a bet of 1/2 pot it now defends AQs-ATs, 77 by calling, and raises 77 50% of the time. No bluffs to go with 77 in its raising range.
On the T
turn Snowie now checks it's 77/AT sometimes, and bets the rest of it's range at 1/4 pot. When the BU faces this bet it still recommends raising with a range of only bluffs, 39% of the time now.When BU calls and the river is the 4
, Snowie now recommends a mixed strategy with all hands except 77 which is always checked. It still leads with the 1/4 pot sizing. Now when Snowie gets shipped on (still all value) it wants to fold 80% of the time (opponent risking 75.25 to win 64 in the scenario I set up)Snowie thinks BU should call twice with hands like KT instead of turning them into bluffs, despite them not beating anything in UTG's betting range.
Still some wacky stuff...