Thank you
Statistics: Posted by Timmy1992 — Tue Oct 17, 2023 6:18 pm
]]>
For example, if you have AA and the flop comes AT2 rainbow, what’s the thought process behind the recommended bet, would it’s bet recommendation be based on getting most value across all streets, including the streets yet to come, or just winning the hand on the street it’s on. Does it understand building a big pot because you have a huge hand? I hope that makes sense.
Thank you.
Statistics: Posted by MetZo94 — Sun Aug 13, 2023 7:18 pm
]]>
]]>
My understanding of indifference is that in a Nash Equilibrium strategy every action at some decision point with some hand will result in the same EV, unless the probability of taking that action according to the NE strategy is one or zero.
Given that definition, I don't see why the difference in EVs between different actions at different decision points would always be the same as I think you are suggesting in your shove fold example
I don't follow your Kuhn poker example, but https://skranz.github.io/gtree/articles ... poker.html seems to support my definition of indifference.
Of course, it must always be the case that a player is indifferent between all moves over which he mixes in an equilibrium.
So you would expect to see the same EV when the strategy is mixed, but different EVs when one action should always be taken and the other never.
Statistics: Posted by spears — Thu Jun 01, 2023 7:02 am
]]>
1. Change in stack size relative to current decision point (this gives the ubiquitous EV(fold) = 0 ).
2. Change in stack size relative to beginning of hand (EV(fold) = final stack - starting stack ).
3. Stack size at end of hand (EV(fold) = final stack ).
All of these referece points will produce different EVs.
But if we stick to one reference point for a given decision point, and compare the EVs of all the possible actions, the difference between EVs will remain the same.
That is a confusing sentence so an example will help:
HU $0.5/$1 NLHE
Starting stack sizes of SB and BB = $10
SB posts $0.5
BB posts $1
SB first to act with AA, has 2 options fold or shove.
BB calls a shove 30% of the time.
BB's calling range has 20% equity against AA.
Reference 1:
EV(fold) = $0 EV(shove) = $3
Reference 2:
EV(fold) = -$0.50 EV(shove) = $2.50
Reference 3:
EV(fold) = $9.50 EV(shove) = $12.50
When comparing the difference in EVs within a reference point, shoving is always expected to make $3 more than folding.
What confuses me is when I try to apply these different reference points to the indifference principle in a Kuhn poker situation.
https://en.wikipedia.org/wiki/Kuhn_poker
Lets assume both p1 and p2 start with stacks s, then they both ante 1 chip.
Lets assume p1 always bets (1 chip) with K, always checks Q, and sometimes bluffs (1 chip) with J.
When p2 is deciding how often to bluff with a J after p1 checks, he wants to make p1 indifferent to folding and calling with bluff catchers (Q).
EV(p1 folds Q) = EV(p1 calls Q)
Reference 1:
0 = P(p2 has J)*P(p2 bets J)*(3) + P(p2 has K)*P(p2 bets K)*(-1)
0 = (1/2)*P(p2 bets J)*(3) + (1/2)*(1)*(-1)
0 = (3/2)*P(p2 bets J) - 1/2
P(p2 bets J) = 1/3
Reference 2:
EV(p1 folds Q) = (s-1) - s = -1
EV(p1 calls Q) = P(p2 has J)*P(p2 bets J)*(s + 2 - s) + P(p2 has K)*P(p2 bets K)*(s - 2 - s)
-1 = (1/2)*P(p2 bets J)*(2) + (1/2)*(1)*(-2)
-1 = P(p2 bets J) - 1
P(p2 bets J) = 0!?
Reference 3:
EV(p1 folds Q) = s-1
EV(p1 calls Q) = P(p2 has J)*P(p2 bets J)*(s + 2) + P(p2 has K)*P(p2 bets K)*(s - 2)
s-1 = (1/2)*P(p2 bets J)*(s+2) + (1/2)*(1)*(s-2)
P(p2 bets J) = s / (s+2) !?
Why are all of these answers different if I'm sticking to the same reference point throughout the calculation?
Statistics: Posted by namnipeels — Tue May 30, 2023 9:31 pm
]]>
Statistics: Posted by wasemunck — Wed Apr 26, 2023 7:37 pm
]]>
At the beginning of the game both players are assumed to have an equal probability of holding all the hands that are not excluded by the opponent's actual hand.
At the beginning of a subgame both players are assumed to be holding some distribution, but apart from that the approach is the same as playing the entire game.
If you haven't done so already take a look at Noam Browns NIPs 17 lecture on YouTube
Statistics: Posted by spears — Thu Feb 02, 2023 6:24 pm
]]>
The core communication is through web requests sending information back and forth with not real way to bury deep enough to monitor what you do within the mobile or desktop environment. Even if that were the case, virtual machines would make it trivial to bypass.
Hiding bot behavior from poker sites will be limited to speed and frequency of certain decisions and how often decisions deviate from optimal.
This concept applies to almost all technologies with server client architecture. Games would be the minority in how extensive their reach is into a system through the help of anti cheat software.
Hiding almost anything in this realm is handled pretty much the same way. Make it look human as possible. Provide randomness of choices and data as well as timing differences in frequency. It’s how hackers bypass modern day anti malware and firewalls the same way you would work automating a casino client. It’s not rocket science.
Statistics: Posted by munawir — Thu Feb 02, 2023 12:57 pm
]]>
It seems to me that it's impossible to pick one strategy for poker that will work all the time, because every time the game and the players are different. A few months ago I was playing poker, but soon I got tired of losing because I couldn't win with one strategy.
That's the point of Nash Equilibrium solutions. They wont lose against any strategy (provided you play for long enough)
Statistics: Posted by spears — Thu Feb 02, 2023 12:02 pm
]]>
]]>
]]>
]]>
I am currently stuck at the understanding of unsafe subgame solving. I understand that each player is maintaining a opponent hand distribution. But how do I build the subgame to solve or how do I use these distributions?
For example sample a pair of cards for the opponent and take the ai cards and maybe already revealed public cards to do CFR with?
Does someone knows how unsafe subgame solving works in detail?
Thanks
Statistics: Posted by lauchi — Sun Jan 22, 2023 11:37 pm
]]>
]]>
]]>
Eg solve for flop 2c3c4c -> regret sum and strategy sum is passed to 2c3c5 solve as a starting point -> solve for 2c3c5c.
This seems like a good way to speed up solving over a range of flops? Is this commonly done? Has this been written about in any research papers anywhere? or maybe I'm missing something and this wouldn't work at all? gimme any input u got
Statistics: Posted by b4rny — Sat Oct 22, 2022 3:19 pm
]]>
Thanks, waiting Dm
Statistics: Posted by xPandaX — Mon Aug 22, 2022 9:46 pm
]]>
]]>
]]>
distribution aware for flop&turn and ochs for river (or just win and draw% if doing a very coarse abstraction)
potential aware is a little better than distribution aware, but harder to implement
clustering algorithm is just kmeans
Statistics: Posted by llllllllll — Wed Jun 29, 2022 4:45 pm
]]>
does anyone have any insight or code into the exact current methodologies for creating flop/turn/river buckets?
I'm not up to date with this stuff, but I'd say read Noam Brown's papers and if it's not clear follow up his references.
Statistics: Posted by spears — Tue Jun 21, 2022 8:46 am
]]>
]]>
]]>
]]>
Create a strategy with a bunch of parameters, let's say a 1000. The strategy would be designed in a way that some choice of parameters would result in a strategy that is "good", i.e. not very exploitable.
For example, there could be a bunch of rules such as - on the flop in position, value bet hands with p_1 % equity against opponent's range and bluff p_2 % of our range.
But the question is, how to optimize the parameters, ideally via some CFR-like algorithm?
Statistics: Posted by listerofsmeg — Sun Apr 17, 2022 6:19 pm
]]>
It should be able to play 6-max and 9-max tables, shouldn't require a lot of computation to act and it would be nice if it could be adjusted in some way to exploit the environment.
What is a good approach?
I've been following the poker research for years but this still seems like a somewhat uncharted territory. The state-of-the art seems to be Pluribus, but it's very complex, it's 6-max, requires fixed starting stacks and uses a lot of computation resources. And the near-GTO strategy may not even be appropriate in a highly exploitable setting.
Right now I'm thinking about this, although it's a very vague idea - create some sort of a parametrized strategy template with hundreds or thousands of parameters and optimize the parameters via self-play.
What do you think?
Statistics: Posted by listerofsmeg — Sat Apr 09, 2022 7:36 pm
]]>
/* Generates legal suit patterns for all legal rank patterns.
*
* Legal rank patterns are 222, 223, 233, 234 but not 232, 235 or 333. For speed only prototypes are generated.
* Legal suit patterns are ccc, ccd, cdc, cdd, cdh but not cch, dcc.
* Not all legal suit patterns can be combined with all legal rank patterns. Clearly 222 and ccc are impossible.
* Some rank suit combinations are duplicates. eg 2234 cdcd and cddc
*
* The duplicates are removed finding ranks by suit. eg 2c2d3c4d becomes c23 d24, and 2c2d3d4c becomes c24 d23.
*
*/
Strikes me the state machine that underlies the evaluator I posted does the job too.
Sorry I can't help more
Statistics: Posted by spears — Fri Apr 08, 2022 9:48 am
]]>
]]>
]]>
Statistics: Posted by dalada17 — Sat Feb 26, 2022 2:10 pm
]]>
basically deepstack failed to beat slumbot but with a few tweaks its performance was greatly improved.
also i'm goint to rewrite big chunks of your code as it's a bit messy(especially nn parts) and i believe performance could be improved. also i found some bugs, would love to discuss more, drop me a message
Statistics: Posted by yegor — Wed Feb 09, 2022 2:15 pm
]]>
It includes pre-trained models and pre-calculated bucketing data, so can be used for testing right away. I did not do any changes to the bucketing or bet sizing approach, but I am looking forward for input how this could be improved...
Statistics: Posted by lucky72s — Tue Dec 07, 2021 4:36 pm
]]>
Statistics: Posted by lucky72s — Tue Dec 07, 2021 4:31 pm
]]>
fully end-to-end, from training to real game play + open holdem connectors.
[email protected]
Statistics: Posted by wolowolo — Tue Nov 23, 2021 11:14 am
]]>
Statistics: Posted by Hissveome — Mon Nov 22, 2021 9:24 am
]]>
project homepage: https://bupticybee.github.io/texassolver_page
github page: https://github.com/bupticybee/TexasSolver
Features:
Completely opensourced
In a tree with 1~2bets + allin, it's speed exceeds piosolver on flop
Support Mac, Linux and Windows
Support texas holdem and shortdeck
Support cross language calls
Result aliged with piosolver
Support dump strategy to json file
It's the c++ version of TexasHoldemSolverJava with a ton of optimization, it's 5x faster than the java version and takes less than 1/3 memory.
Download and try it!
Statistics: Posted by bupticybee — Mon Nov 08, 2021 12:56 am
]]>
I'm looking for several options:
- bot based on something like this http://www.poker-ai.org/phpbb/viewtopic.php?f=26&t=3316
- collaboration for making strong stealth bot (PokerStars, PartyPoker)
- taking part in software development for making such bot (20+ years software engineering: C#, C++, SQL and many others)
Contact me https://t.me/baddy002
Thanks for any assistance
Statistics: Posted by baddy — Sat Oct 02, 2021 6:28 am
]]>
looking for PPP deals on OFC and NLH.
Could you be so kind to clarify for newbie what it means?
Statistics: Posted by baddy — Sat Oct 02, 2021 5:39 am
]]>
Statistics: Posted by Salem — Mon Sep 27, 2021 11:35 pm
]]>
when call the solver using command line tools from Windows, is it possible not to display the console with the log?
Of course it's possible, you can use another program to create a subprocess and run the solving process there. But by default, the console will exist, I mean it's a "console version" right? What kind of console version will it be if it runs without a console?
Statistics: Posted by bupticybee — Sun Sep 26, 2021 3:54 am
]]>
]]>
Hi. Cool job! Is it possible to download the entire source project and open it in Visual Studio 2019 and debug and change the code?
I wrote it using Clion, but at least two other people compiled it under Visio Studio, so I guess yes
Statistics: Posted by bupticybee — Wed Sep 22, 2021 7:27 am
]]>
]]>
Hi, what about multyway and preflop solver?
That's another story, I'm still struggling to make it faster on 2-player game, and a gui, multyway and preflop maybe after that.
Statistics: Posted by bupticybee — Mon Aug 30, 2021 3:24 am
]]>
]]>
Statistics: Posted by bupticybee — Wed Aug 25, 2021 6:10 am
]]>
but it's 5x faster and takes less than 1/3 memory than the java version, and best of all, it's now (under smaller tree settings) faster than piosolver.
check the speed benchmark https://github.com/bupticybee/TexasSolver#speed-benchmark-with-piosolver i have done with piosolver.
| | Input config | thread number | memory usage | accuracy | converge time |
|:------------------------------:|:------------------------------:|:-------------:|:------------:|:--------:|:-------------:|
| piosolver 1.0 | piosolver 1.0 | 6 | 492Mb | 0.29% | 242s |
| TexasSolver 0.1.0 (Our solver) | TexasSolver 0.1.0 (Our solver) | 6 | 1600Mb | 0.288% | 172s |
So why spent 1000$ on piosolver? TexasSolver is free and opensourced.
Statistics: Posted by bupticybee — Mon Aug 23, 2021 3:04 am
]]>
So far I came up with 3 ideas:
1. Instead of storing the HS as a float I can use unsigned chars, it won't be as accurate but it will be much smapper. The logic is to assign an 8 bit char between 0 and 255 to a HS value between 0.0 and 1.0, so 0 = 0.0, 255 = 1.0, 122 = 0.5 etc.
2. I could write my own isomorphism algorithm, this could be more space efficient than K. Waugh's implementation because I could collapse hands that don't make flushes. However this is a pretty hard thing to do and I don't know how much space I would really save, especially since this would be my first time writing an algorithm like that.
3. Instead of using HS on the river, I can use some sort of absolute hand strength. For example 10 buckets for straights, 10 buckets for flushes etc. I think this will not be as effective as using HS2, but it would be much easier to compute, since I have to evaluate the hands anyway during the CFRM training.
monker does #2:
Omaha postflop
Omaha is too large to solve completely even when starting from the flop. MonkerSolver therefore reduces the game size by making the following concessions:
Before the river, only suits which has the potential to become flushes are considered. For example, these two hands are merged, even though they influence the opponents range slightly differently:
A♠K♠J♠5♠ 6♣5♣2♣7♣
A♠K♥J♥5♦ 6♣5♣2♣7♣
On the river, only suits with 3 or more cards on the board are considered. For example, consider these hands:
A♠K♠A♦K♦ 6♠5♠2♥7♥T♥
A♣K♣A♦K♦ 6♠5♠2♥7♥T♥
The above two hands are merged into the same bucket. They are however not strategically equivalent, since we are more likely to be up against a missed flush draw with the latter hand.
Statistics: Posted by llllllllll — Sat Aug 21, 2021 5:16 pm
]]>
It's your standard river solver with a betsize solver attached, which seems to converge to a stable set of optimal betsizes at each node even in deeper trees after applying a novel bet size gradient.
Still some bugs, but the example in setup works for a simple tree. I'm working on fixing these bugs and eventually on creating a website where you can play around with the solver. In the meantime, feel free to check out the code, test some examples, and ask questions.
I'm pretty new to programming, and this is my first serious project I've taken on. So I appreciate any tips or advice you might want to share.
Statistics: Posted by churchkleene — Sat Jul 31, 2021 1:53 am
]]>
]]>