u1 = 1 and u2 = 0 during the implementation of the above
algorithm. Thus $1 is always moving and eventually oscillates
around the mean vector of the two clusters of samples, while
dz is always the loser and remains fixed during the whole
learning process. Such a loser unit is usually called a dead unit.
Unfortunately, the situation of dead units happens quite
often for the above simple CL algorithm unless the weight
638 IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 4, NO. 4, JULY 1993
vectors are initialized near the cluster centers which are the
eventual points of attraction. As mentioned previously in
Section I, many efforts have been made to solve the problem.
An often used strategy, sometimes called conscience, is to
reduce the winning rate of the frequent winners. For example,
in Fig. l(b), when GI initialized at A2 wins a certain number
of times, we force u1 = 0 and u2 = 1 to bring G2 according
to (lb) towards its point of attraction. By doing so, G2 will
be gradually brought towards the right-side cluster while GI
keeps oscillating between the two clusters. Finally W; will
stop oscillating and the two weight vectors will converge to
the two cluster centers. The so called FSCL
u1 = 1 and u2 = 0 during the implementation of the abovealgorithm. Thus $1 is always moving and eventually oscillatesaround the mean vector of the two clusters of samples, whiledz is always the loser and remains fixed during the wholelearning process. Such a loser unit is usually called a dead unit.Unfortunately, the situation of dead units happens quiteoften for the above simple CL algorithm unless the weight638 IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 4, NO. 4, JULY 1993vectors are initialized near the cluster centers which are theeventual points of attraction. As mentioned previously inSection I, many efforts have been made to solve the problem.An often used strategy, sometimes called conscience, is toreduce the winning rate of the frequent winners. For example,in Fig. l(b), when GI initialized at A2 wins a certain numberof times, we force u1 = 0 and u2 = 1 to bring G2 accordingto (lb) towards its point of attraction. By doing so, G2 willbe gradually brought towards the right-side cluster while GIkeeps oscillating between the two clusters. Finally W; willstop oscillating and the two weight vectors will converge tothe two cluster centers. The so called FSCL
การแปล กรุณารอสักครู่..
