search

xurxodiz / MarioLevels

My entry for the Mario AI Championship 2012, implementing adaptive procedural level generation. Holds 1st place at the all-time points and percentage tally.
MIT License
2 stars 1 forks source link

Pick a clustering algorithm #23

Closed xurxodiz closed 12 years ago

xurxodiz commented 12 years ago

To create clusters for the data gathered in #22. Héctor recommends Expectation-Maximization as done by Weka. Is it our best choice?

We'll be giving it a couple dozens of entries, each with around thirty variables, and we expect to get the clusters, the entry-cluster associations and the centroids from it.

xurxodiz commented 12 years ago

Currently examining the different results in Weka. It looks like any of the available algorithms creates two clusters, except for EM that organizes the data in three.

xurxodiz commented 12 years ago

EM now finds four clusters, but one of them is pretty small. Intuitively, best results are obtained in all algorithms for three clusters.

And out of them, the best results (clearer separation and definition of each cluster) are given by EM and XMeans; but, again, much better for EM.

So we're gonna roll with that.