View Single Post
  #1  
Old 06-08-2012, 10:10 AM
markweitzman markweitzman is offline
Invited Guest
 
Join Date: Apr 2012
Location: Las Vegas
Posts: 69
Default Question 14 LLoyd's algorithm - empty clusters

While checking my implementation of LLoyd's algorithm, I received an error of float division by zero on one of my runs (it seems to happen rarely). I realized that if on the initial or subsequent iteration one of the clusters has zero elements then the update for mu for that clustered needs to be altered (see slide 11/20 Draft Lecture 16 for update rule). There seem to be two possibilities - either use the old value of mu for the cluster (i.e. not update it) or seed a new random value. If you use the old value then it is possible that the cluster will never get any members (for example if your points have a relatively empty region in the square and the mu for the cluster is in that region. Effectively you are using one less cluster if that happens. The alternative is to use a new random seed for that cluster and hopefully get out of the empty region but I am worried about convergence issues.

I am curious if others have encountered this issue and how they think it should be dealt with.
Reply With Quote