Poker-AI.org

Hi,
can you tell me if this is right for earth movers distance?

Seems to work.

Can you elaborate on those graphs?

I clustered all flops using KMeans++, ehs2-histograms with 100 bars, and EarthMoversDistance. One graph = one cluster. One line = one flop. The heat-pixels are the histogram-bars. I hope that was understandable.
Did this (for testing purposes) with k=200 and only 10 kMeans iterations.

Your goal here is to cluster public board buckets, right? Did you use all possible hole cards to create the EHS2 distributions? And if so, did you check the boards within a cluster if they are "meaningful", i.e. represent boards that one would typically group together? I tried this too, but did not get the results I expected.

Goal here was to cluster complete flops (= 2 hole cards + 3 board cards). The ehs2-distributions are against all possible holecards. I created the graphs to evaluate the results. But they are not as good as they look at first view. Some hands with totally different distributions are clustered together. I believe (and hope) that is because I did not use enough clusters and ran only very few iterations.

For everyone that experimenting with k-means with the Earth Mover Distance, are you guys using the traditional average to compute the centroids of each group (that is treat each histogram as a vector and compute the average of all vectors in a group)? I ask because the average is not guaranteed to be the centroid of a group for metrics other than euclidean. Do you guys think that is what these papers did? Maybe the average is good enough...

I was so confused by this I asked a question on stack exchange: http://stats.stackexchange.com/question ... 823#112823. Someone there provided a good example of why averaging does not always give you the best centroid. Any comments on this?

Hi Guyz,

I have decided to code a bit and give a try to this method.

I have clustered preflop hands in 8 clusters and these are the results :

Not exactly the same as in the paper...

My histograms aren't really the same as well :


What are your results for preflop clustering and histograms ? Do they match the paper ?

Regards,

MrNice

If I calculate the EMD of all possible histograms in the first round of the game, I get a 169x169 matrix like the small 4x4 EMD matrix in Figure 2 (page 4) of the paper.
After performing kmeans clustering on the 169x169 matrix (kmeans++, multiple restarts, 8 clusters) I get a slightly different result like in Table 1 (page 5) of the paper.
My histograms and EMD values are the same like in the paper that is why I think I am doing something wrong between EMD calculation and clustering.

I think they reduce the 169x169 EMD matrix to a 169x1 EMD matrix because otherwise the matrices are getting pretty huge on the second and third round of the game, right?
I am happy for any hint and I am also interested in exchanging information with others who are on the same journey .

I have used the traditional average in the past. I know about the theoretical concerns, but for me it still worked fine. The clusters look reasonable and the resulting bot was pretty good.