Fig. 1. A crucial problem for competitive learning. (a) The desired result of
using CL: each of weight vectors moved to a cluster center regardless of the
initial values of these weight vectors. (b) The actual result highly depends on
the initial weight vectors. W ' l , t & will move to cluster centers when they are
initialized at points A I ,B1, respectively; but when W'l is initialized at A2
and 122 at B2, $1 will be located around the middle point of the two clusters,
while Gz remains unmoved as a dead unit. (c) A typical result of using the
conscience strategy and FSCL when the number k of used units is larger than
the number of clusters in a data set, where the two extra units have also been
moved to some boundary points between different clusters, and may confuse
quite a large portion of samples from different clusters.