Information Technology Reference
In-Depth Information
0.1
0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
0
500
1000 1500
Iterations
2000
2500
Figure 2.8 Deviation from the MC direction for LUO, measured as the squared sine of the
angle between the weight vector and the MC direction.
30
LUO
LUO ODE
25
20
15
10
5
0
0
500
1000
1500
2000
2500
Iterations
Figure 2.9 Check of the validity of the ODE approximation for LUO.
2.6.2.6 Stop Criteria It is impossible to use the usual stop criteria for the
MCA learning laws, because the error cost does not go to zero at convergence.
Indeed, the minimum of the MCA error cost is equal to the smallest eigenvalue
of the autocorrelation matrix. Hence, it is necessary to detect the flatness of
the curves representing either the weight components or the error function. This
problem is complicated by the difficulties described in Proposition 69. So this
type of detection is valid only for MCA EXIN, which has the slowest divergence.
Search WWH ::




Custom Search