158
T.-S. Chon . Y.S. Park· I.-S. Kwak . E.Y. Cha
training set, while data collected from April 1997 to March 1998 were used as new
data for testing the trained network.
When community data were trained with the RTRN, convergence was usually
reached between the 5,000th and the 1O,000th iteration. The training sets were well
in accord with the matching output. In order to verify the predictability of the
trained network, we further provided new community data from April 1997 to
March 1998. Figs. 8.16a and 8.16b show examples of comparing field data with
the predicted data in different seasons. Generally, dominant taxa such as
Oligochaeta, Chironomus, and Chironomidae showed good matches between the
predictions and the field observations. Pearson's correlation coefficients (Zar,
1984) between the predicted data and the field data ranged from 0.55 (F=34,
P
0.4 on these types of community data (Chon et al. 2000b).
8.3.4
Impact of Environmental Factors Trained with the Recurrent Network
Revealing the impact of environment on communities is essential for finding
causality of community response to disturbances. The impact of environment
could be judged by prediction of community abundance corresponding to
variations of environmental factors, and many researches have been conducted on
measuring the impact of environments in community based on the sensitivity
analysis (e.g., Dimopoulos et al. 1995; Scardi 2000; Recknagel and Wilson 2000;
Salvador et al., 2001). In this case, however, the time setting has been usually
static.
As mentioned previously, however, the time factor is one of the key issues in
community dynamics especially in the context of regressive or progressive
community changes. In this case, in order to emphasize time-dependency in
community data, we trained changes in community and environment in the scheme
of recurrent training. The previously mentioned the fully connected recurrent
network, RTRN, was modified to accommodate environmental factors, but, unlike
the community data, neurons accepting environmental factors did not have
recurrence feedbacks (Fig. 8.15b). In concurrence with the input of biological
data, the corresponding sets of environmental data were given to the modified
RTRN, producing, through the connectivity of the network, continuous,
independent effects on determining community abundance. In addition to 7
neurons for community data for external inputs and 13 neurons for the hidden
layer, 4 neurons were used for receiving environmental factors separately.
As for environmental data (E), monthly observations of water velocity and
depth, amount of sedimented organic matter, and volume of substrates smaller
than 0.5 mm were provided to the network. The relationships between the
environmental data and community dynamics were successfully extracted, and
predictability was greatly increased when the data were trained during aperiod of
strong environmental influences Ce.g., flooding). The predicted data by the
network trained with the community plus environmental data were distinctively
T.-S. Chon . Y.S. Park· I.-S. Kwak . E.Y. Cha
training set, while data collected from April 1997 to March 1998 were used as new
data for testing the trained network.
When community data were trained with the RTRN, convergence was usually
reached between the 5,000th and the 1O,000th iteration. The training sets were well
in accord with the matching output. In order to verify the predictability of the
trained network, we further provided new community data from April 1997 to
March 1998. Figs. 8.16a and 8.16b show examples of comparing field data with
the predicted data in different seasons. Generally, dominant taxa such as
Oligochaeta, Chironomus, and Chironomidae showed good matches between the
predictions and the field observations. Pearson's correlation coefficients (Zar,
1984) between the predicted data and the field data ranged from 0.55 (F=34,
P
0.4 on these types of community data (Chon et al. 2000b).
8.3.4
Impact of Environmental Factors Trained with the Recurrent Network
Revealing the impact of environment on communities is essential for finding
causality of community response to disturbances. The impact of environment
could be judged by prediction of community abundance corresponding to
variations of environmental factors, and many researches have been conducted on
measuring the impact of environments in community based on the sensitivity
analysis (e.g., Dimopoulos et al. 1995; Scardi 2000; Recknagel and Wilson 2000;
Salvador et al., 2001). In this case, however, the time setting has been usually
static.
As mentioned previously, however, the time factor is one of the key issues in
community dynamics especially in the context of regressive or progressive
community changes. In this case, in order to emphasize time-dependency in
community data, we trained changes in community and environment in the scheme
of recurrent training. The previously mentioned the fully connected recurrent
network, RTRN, was modified to accommodate environmental factors, but, unlike
the community data, neurons accepting environmental factors did not have
recurrence feedbacks (Fig. 8.15b). In concurrence with the input of biological
data, the corresponding sets of environmental data were given to the modified
RTRN, producing, through the connectivity of the network, continuous,
independent effects on determining community abundance. In addition to 7
neurons for community data for external inputs and 13 neurons for the hidden
layer, 4 neurons were used for receiving environmental factors separately.
As for environmental data (E), monthly observations of water velocity and
depth, amount of sedimented organic matter, and volume of substrates smaller
than 0.5 mm were provided to the network. The relationships between the
environmental data and community dynamics were successfully extracted, and
predictability was greatly increased when the data were trained during aperiod of
strong environmental influences Ce.g., flooding). The predicted data by the
network trained with the community plus environmental data were distinctively