en stronger.
The implementation of the above algorithm is quite easy;
each weight vector Gi is randomly initialized, and then the
above two steps are iterated until the iteration converges or
freezes as the learning rate a, becomes zero or very small, or
until the number of iterations reaches a prespecified value.
For a data set D consisting of several clusters of samples
(e.g., three clusters as shown in Fig. l(a)), when the number
of units k is equal to the number of these clusters, the desired
results of implementing the above algorithm should be that the
weight vector of each unit is moved to the center of one of
the clusters (Fig. l(a)) regardless of the initial values of these
weight vectors. However, the actual results highly depend on
the initial weight vectors. For example, even for a two clusters
problem shown in Fig. l(b), when the two weight vectors
GI,G2 are initialized at positions A I ,B1, respectively, they
finally move to the center points of the two clusters after the
learning has converged. However, when dl is initialized at
position A2 and G2 at B2, the result of (1) will always be