Poker-AI.org

Hi everyone,

First of all, thank you for this amazing forum ! I am new in the hobby of poker botting and find all your resources incredibly helpful I am building my first poker AI, building on Open Pure CFR implementation, read every paper I could found on the subject, and find it really fascinating.

However, one thing that strikes me is that all the litterature seems to consider the stack size as a constant variable. While this is approximately true for cash game if players are using 'auto-rebuy' options (i.e if I have 100$ and I lose 10$ on a specific hand, the poker software direclty completes my stack to 100$), tournament play means being able to play at different stack levels, and the strategy are completely different when the stack levels vary. For instance, depending if we are 10bb or 200bb deep, I would rather have K5o or 67s.

Because I want to test my bot on Hyper Turbo HU SNG (with 25bb to start with), I was tempted to learn 25 different strategies and use soft geometric translation for each hand in order to determine which strategy I will be using (for instance is my stack size is 24.3bb I would use the 24bb strategy with probability 30% and the 25bb strategy with probability 70%). However, it seems to me that :

There should be more memory-efficient & time-efficient way to learn a global strategy for tournament play.

At least, I would like to somehow use my 24bb strategy as a warm start for learning the 25bb strategy.

Has any of you give any thought about this and/ or are you aware of any scientific work tackling this issue ?

Yes I've pondered this for a few years but haven't got very far.

The only academic work I can think of that is restartable is https://www.cs.cmu.edu/~ggordon/mcmahan ... bundle.pdf

Thanks a lot ! I am going to give this paper a closer look and let you know if I come up with some valuable idea in the context of CFR solving.