Poker-AI.org

I want to train my neural networks to predict the opponents move for a given situation. One vital information of course are the pocket cards the opponent holds.

I have a database of hand histories an it's easy to assign the cards when a specific hand history reveals the opponents pocket cards (showdown or just because the opp shows his cards).
But this of course is the minority of the histories.

How can I predict the hole cards when there are no informations about it?

I have 169 pre flop buckets. I thought maybe I even out the probability of playing those buckets and when the opp reveals his hole cards I add some weight to this bucket.
After doing this for each history I have a weighted list of buckets and over this I could do two things:
Always pick the most likely hand that he will play, which seems not very clever.
Or I could pick a random pair of hole cards based on the weights. But this approach seems pretty random to me.

I thought that I maybe could build some relations between the actions the player takes on the hand and his hole cards, but this still leaves most of the hands out (like for example pre-flop folds or flop folds).

Has anybody already dig into this problem? It seems like a pretty vital part to me to successfully train a neural network, but I couldn't really find any information on this matter.

I fight with similar problems.
Atm I do not predict cards, but Hand Strength and Hand Potential. That seems much easier than predicting the exact hand. From the knowledge you gain from showdown hands you can try to learn something about the hands without known cards since you have some info about that distribution. It is some kind of voodoo, I did not get very far with it. The literature about it is not very precise about the results they get. And some papers do not explain very well how they do what they do (e.g. exact input into the NN).

Atm I'm thinking about first filtering the hands that are a clear bluff, to have a more exact prediction of the non-bluff strength...

Edit: not to steal the thread, but a paper with too few detail for a noob like me is the following. I get the basic idea, but I'm too dumb to apply it from the details they give: Using Kullback-Leibler Divergence to Model Opponents in Poker