Robuste Abstraktion für Heads Up No Limit Texas Hold’em Poker

* Author: Thorsten Spieker
* Department: Cognitive Systems Group
* University: Otto-Friedrich Universität Bamberg
* Language: German

Abstract:
New methods for successfully abstracting the action sequence tree of heads up no limit hold'em are presented.
- Restriction of total number of raises allowed instead of restricting number of raises per stage
- Finding optimal betting size abstractions by analysing a big database of hand histories
- Introducing a new method of comparing the quality of an abstraction without the need to create and let a strategy play.

Link:
http://www.cogsys.wiai.uni-bamberg.de/t ... tPoker.pdf

Looks interesting. Author might consider reaching a broader audience with an English translation.

I'm interested in knowing about the quality of the 'abstraction of betting sequences' part,
my german is a bit too weak to understand it but if it's good I might have to try harder...

Key point of that part is that restricting the total number of raises allowed in the game (raises preflop + raises flop + raises turn + raises river) to reduce the size of the game tree is better than restricting the number of raises per street.
It reduces the tree much more and yet covers the game better how it's played by humans.

e.g. if you restrict the total number to 8 then you have a max depth of the tree to 8 (+a bunch of calls) instead of restricting per round to 8 where you then have a max depth of 32 (8+8+8+8+a bunch of calls). The sizes is reduced a lot in the first tree but does it represent the real game enough or even better? Yes it does. For the second version you would allow a game that could have 32 raises ... not gonna happen in real life. For the first version you allow 8 total and can distribute them wherever you like while you would need to allow 8 raises per street if you want to allow the same diversity. Now why 8? Well its an arbitrary number but the key idea is, if there are x number of raises then its very likely that one of the players is allin. It doesnt matter on which street these raises happend. To support this claim there are tables about coverage of a 40million hands database for different abstractions. Basically if you restrict like this you can reduce the tree while increasing the accuracy of the abstraction.

Vielen dank!