Poker-AI.org

I have an idea on how to estimate opponent's actions and hand range.

Given we have a big volume of hand histories, we could try to match hands that are similar.

We would say that hands are similar if players have similar stats, they do similar actions and board texture is similar. Of course I am not talking about exact matches, lots of abstractions would be needed.

Then, from these hands we could predict opponent's feedback to our actions and hand ranges (out of hands with showdown).

This approach is different than common approach with NNs and I haven't read about it anywhere. Is there any reason (ignoring performance) why it wouldn't work? Maybe anyone already tried that?

You are basically assuming every player is the same which is not a valid assumption. For instance, the board is 9h8h4s, you cbet and villain raises. Some players do it only with sets and 98. Some stack off lighter with hands like AA/KK, some have monster draws in their range.
If you want to be exploitive, its crucial to get an idea if the player is for example raising his strong draws or just calling them. If you want to be GTO, let the algorithm find out if it makes sense to raise hands from the different categories.

I'm not saying this couldn't work, but it's definitely a very basic heuristic and there should be better solutions.

proud2bBot,

Maybe I did not describe it accurately. I am not assuming that every player is the same. Instead, I am assuming that there are n categories of players where players in each category are the same, so I am only searching for hands of players in given category.

you can easiliy do it this way and categorize you opponents by stats. Afterwards adding some weighting factors (e.g. you see opponentX calling with small Suited Connectors) and let this influence your predictions to receive a more accurate range.

Thats what I understood, too. The question is: how do you define the category. As I understood it, you are clustering players according to some preflop or postflop stats. And my point was that this doesn't work well, as stats are an aggregate from which you cannot really infer the strategy too well. For instance, Assume preflop and we have 2 players with exact same VPIP/PFR/3b stats. Both 3b 8%, but P1 does it with a polarized range while P2 uses an unpolarized range - you would merge this together, right?

Thanks, I understand your concern now. Yes you are right, the plan is to cluster by stats (didn't choose the concrete stats yet).


As with your example, I'm thinking about it this way: consider these players have the same 3bet, vpip and pfr, but different cbet and donkbet, so they will end up in different clusters. I might be wrong here, but I believe that if we take an appropriate set of stats the resulting clusters shoul have pretty much similarly acting players.



Any of you guys already tried to implement something similar?

yes I have implemented something similar like this. I'm have some cornerstones to categorize the opponent -> So I get the stochastic ranges where I have some factores which will flue into the range and action calculations which has direct impact to my finally decision.

I guess if you take a lot of stats into consideration it might create clusters that mimic players approximately. However, considering a single stat has say 8 partitions and you are using 4 stats only, that will lead to 8*8*8*8=4096 different player types already...

If you take the naive approach you are right but you'll have lots of empty categories that can probably be pruned away in the end.
Personally I haven't done any research in that direction, especially NN's, but I do see the merit in this. Actually this wouldn't be so much of an NN problem but a clustering problem to begin with, then it becomes a prediction problem based on the clusters.

If winnie got this to work already this should definitely have some value.

_________________
Cheers.