Poker-AI.org

Are there any card abstractions / bucketing methods available online that can take a hand on the flop, turn and river and spit out an "equivalence" group. (Almost like a lookup - for implementing in code)

Such that AJ on K73r is roughly equivalent to AQ on K72r to shrink the game tree.

From some of the reading I've done, the HU Limit bots used 169 buckets for preflop into 9000 buckets for flop, 9000 for turn and 9000 for river. Something of that magnitude would be great but I'll settle for any real way of classifying hands for now.

Thanks!

The average and variance hand rank of AJK73 on the river is similar to the avarage and variance hand rank of AQK72 on the river.

Not sure what you mean. I thought most abstractions used E(HS) and E(HS^2) to categorise hands?

E(HS) and E(HS^2) will group too many very different hands together in this scenario
Hand rank is a number that erm... ranks a hand. Royal Straight Flush is top rank
Average and variance arise from rolling out all future cards to river.

I see, so E(HS) is not a great method. I suppose the publications that I read that used it were 10 years old.

How would hand rank be better though? There is no "flush draw" rank for example because it's not related to showdown.

What do you mean by averaging and variance? We enumerate all possible cards and see how our hand rank changes?

I didn't say that. Ehs isn't good for the application you describe. If you think it is why don't you use it?


Flush draw has high variance


Find all possible rivers that can develop from the hand.
Find the hand rank of each one
Find the average and variance of the list of hand ranks

What about this application makes it bad, from my understanding now, I thought it was used 5+ years ago but had been succeeded by other methods that built on it and are superior in every way? (Accuracy + Computation time mainly)



Oh I see what you mean. That does have a lot of merit to it. From the publications I was reading they use a similar method but with earth mover's distance instead (EMD). I don't understand EMD enough to plug it in right now but I'll take a better look at it.

How would you recommend splitting it using your method (perhaps split into x categories by 1/x of the equity and then further split inside with variance? Or the other way around) and would you recommend looking at equity distribution over turn AND river too? The leading bots do that

Things would be a lot easier if someone had done all this already and stuck it on github

I think you have understood the idea now. So you can answer your latest questions yourself with some experiments.

Have you done any work on these things yourself in the past?

Yes

... Are these available anywhere or just your own private work?

Although, honestly, I'm sick of looking at code. I'd rather someone had summarised the theory behind these things in such a way that would make it NOT appropriate for academic work.

If it was easy and publicly available, there would be no money to be made from it

I didn't realise there was much money to made from it tbh