Chapter 8· Analysis of Stream Macroinvertebrate Communities
155
Moutputa
(Communily)
T:timedelay
Moutputs
(Community)
T: time delay
a)
N-M hidden outputs
Proceealng layer
Concatenated layer
Mlnputs
(Communlty)
Minputs
(Community)
Processing layer
Einputs
(Environment)
b)
Fig. 8.15. A diagram of the real-time recurrent network for patterning community
changes in benthic macroinvertebrates in streams. (From Chon et al. 2(01). a)
input data: community. b) input data: community plus environment.
The learning rate and the momentum coefficient were 0.3 and 0.7, respectively. In
this study, 7 neurons were used for community data for external inputs, and 13
neurons for hidden nodes. The error term was the sum of the difference between
the output and the target data for all nodes (selected taxa) for all patterns (sampIe
sites), and the criterion of the error term for allowing convergence was 0.006. The
data for the previous three months were given as the input in a sequence with
recurrent feedback, whiIe the data for the fourth month were provided as the
matching output.