Chapter 8' Analysis of Stream Macroinvertebrate Communities
163
new weights UP+l) and Vjk(t+l) were updated by the iterative process as shown
in the process 4 in Fig. 8.19. Alpha (a.) is a constant determining the learning rate
and the values around 0.3 were used in this study. The detailed process in the
Kohonen network could be referred to Kohonen (1989) and Chon et al. (1996).
1. lnitialize weights for the networks.
2. Present new input.
3. Finrling the best matching processing element on Kohonen layer
I, = 'i Uij(t)xj + i>:.(t)y.
I-I
"'=1
where Xj and y. are input data, and Uij and v,. are the weight
from input node to Kohonen layer.
4. Update Weights on Kohonen layer.
U/t+ I) =Uij(t)+a(xj -Uij(t»Z,
where a is a learning rate of forward flow on Kohonen layer, and
z. - {I if li=li* Vj (i*;winning node)
, - 0 otherwise
5. Present new desired output.
6. Update weights on node p of Grossberg layer.
wl,...(t + I) = wl,...(t)+a(Yk - wl,...(t))Z,
where a is a learning rate of forward flow on Grossberg layer.
(w2qi: similar on node p of Grossberg layer 1.4).
7. Repeat by going to step 2 until the end of input data.
8. Emit modified desired outputs.
N
y.'= LWljk(t)Z,
};l
(x~: similar)
Fig. 8.19. Algorithm of the counterpropagation network. (From Park et al.
2001a).
After the winner node at the layer L3 was determined, the process was proceeded
to the layer L2 and L4 respectively for Y and X, which are Grossberg layers