Poker-AI.org

Hi,
I've developed quite decent postflop solver (for NLHE) and I am wondering what the best approach for approximating preflop is.
I am interested in preflop games with fixed bet sizes (say raise 3bb, 3bet to 10bb, 4bet to 23bb, push).
What I am doing now is to deal 40-50 flops, solve them and apply CFR. This approach, while producing decently looking results, is not theoretically sound as weak hands will never become 100% folds (once in a while out of those 40-50 flops there will be a few which hit say 32o hard and that will cause spike in regret for a fold).
I was thinking of choosing representative subset of all 1755 flops but I am failing at constructing one.
I was also thinking of rounding say 80+% actions to 1. This approach doesn't work for postflop but maybe preflop there is less mixing necessary.

I appreciate any help/pointers to resources in this respect.

I may be able to help. Can you give me a specific scenarios? I could run it in a sim for most raise criteria such as # of bb, % of stack, % of pot, etc. I'd rather we exchange ideas here than PM, if that is ok.

Hi,

Thank you but I am more interested in general techniques than specific results but as easy example let's try HU 20bb cap with fold/minraise/push preflop.
One approach would be to run CFR on all 1755 flops but that would require significant disk space. I tried dealing 40-50 flops, updating regrets, repeat but it seems there is too much variance in what hands the sample flops hit the most. Do you have any thoughts about how to run this kind of thing most efficiently ?

I have only the vaguest understanding of what you are doing. Can you cluster similar flops together? http://www.poker-ai.org/archive/www.pok ... 694#p36694

It's not that easy to clump similar flops together. If I clump J72 with J82 (making J82 a representative) then 77 is going to be way more profitable than 88.

One thing which I think could work is this:
-there are 1755 strategically different flops in Holdem
-choose 100 of them (possibly with different weights) such that preflop solution to a game with only those 100 flops is very similar to the full holdem

To do that you would need to make sure that in those 100 flops every card appears as often as in all 1755, same for flush draws, paired boards, straight heavy boards... if somehow you could make a set of 100 (or 80, or 120) flops like that then running CFR on whole game tree would be quite achievable on standard hardware. Did any of you heard of efforts to construct subset of flops like that ?

Haven't read of anyone doing this, but then I don't understand everything I read. I think it's a good idea.