Quote:
Originally Posted by yaser
This is a valid approach. You can also restart from new initial centers.

I tried
a) Leave points with 0 nearby events alone, and use them as cluster centers.
b) When there was a point with 0 nearby events, restart from new initial centers.
c) Do 40 random initializations of Lloyd's algorithm and pick the one with the lowest minimum. That always (10000 trials) turned out to have no unoccupied cluster centers.
For K=12 method c) was best at preventing regular RBF from being worse than SVM (problem 16), and for K=9 b) gave a somewhat higher probability of E_in =0 for regular RBF (problem 19). But the three methods were so close together in results that they all gave the same best choice among the multiple choices for problems 1519.