viewtopic.php?f=25&t=2381, but it didn't seem to get any visability so I'll ask it here. A common method it seems to create an abstraction of hold'em is to take all possible hands in a given round (say the flop), and cluster them somehow. The paper discussed in the linked thread is one example of this. For each possible hole-flop combo they compute the distribution of the hand value (that is percentage of possible opponent hands beaten) on the river starting from that flop, and use that distribution as a representation of that hand. The paper then said it used k-means clustering with the earth mover distance (EMD) to cluster the hands. This part was confusing for me because there's no theoretical guarantee that averaging a set of points produces the best centroid for that group for metrics other than the euclidean metric. So is this really what these papers did, or did they use other methods for dealing with this problem (e.g. k-mediods)?
It does seem plausible that using the mean with the EMD should still give decent results, maybe it just works in practice. I went ahead and tried this and grouped preflop hands using this method and it seemed to work ok, but I'm still skeptical that this is what these papers do. Does anyone know more about this?Statistics: Posted by ScoobySnacks — Tue Aug 26, 2014 5:20 am
]]>