Poker-AI.org

As part of a sampling method when comparing strategies, I need about 1000 hands that will represent the entire game space. So, I'd like them to be as homogenous as possible. The 6000 hands used by the U of A during their ACPC(s) looked pretty evenly distributed. Anybody know how they come up with those each year?

i believe they are random, i remember reading their code for dealing cards and modifying it to deal x games

Hmm, I could have sworn they weren't completely random after looking at the HS/HP of the hands. But, that was years ago, maybe I was mistaken. Regardless, any suggestions on generating 1000 hands that represent points evenly distributed throughout the game space?

Are you looking for something like this? https://en.wikipedia.org/wiki/Sobol_sequence

I don't think so.

Another way of approaching this, and applicable to other areas, would be: say you have a ton of data samples that you want to use to train a model. How do you make those samples homogenous as possible across all point dimensions (attributes) as to prevent a biased model (one that favors the most abundant data)? For example, in poker, if you represent your hands via HS/HP metrics, the majority of those sampled will be towards the center of a bell curve.

I guess the obvious answer is clustering. Are there any other ways? I'm trying a reduction method, where I look at the least abundant data and make all the other sample counts equal to that. However, it's only based on a couple attributes.

http://en.wikipedia.org/wiki/Multivaria ... estimation

I think the answer to this question depends on the "training" you are trying to do. Is it for density estimation, regression, classification or something else?

It actually applies to two different problems I'm working on. One is regression, the other is as mentioned above where I'm just trying to represent the entire game space.

"homogenising" the data will invalidate any regression. The only thing I think you can do is increase the resolution of the regression. ie increase the number of neurons, support vectors. Or maybe use random forests.
I'd like to see a toy problem exhibiting this behaviour so I can do some experiments.

You're not really "homogenising" the data, you're picking homogenous points in the game space to sample your data.