168
T.-S. Chon . Y.S. Park· I.-S. Kwak . E.Y. Cha
et al. 2001b). Data for benthic maeroinvertebrate eommunities (Fig. 8.22a) were
provided for ealeulating exergy, and, aeeording to JjIlrgensen (1997), values
ranging from 29.6 - 43.9 were assigned as W; for maeroinvertebrates. In available
information sourees (JjIlrgensen 1997; Fonseca et al. 2000), however, only values
of W; at higher taxa are available, and the values for the order or family level in
benthic maeroinvertebrates are not specifically provided. Based on experienees
from data analyses and field experiences we arbitrarily assigned 30 for
Oligoehaeta, Diptera, Chironomidae and Hirudinae, and 35 for Gastropoda,
Ephemeroptera, Pleeoptera, Trichoptera, Odonata, and Megoloptera in this study
(Park et al., 2001b).
Training between Community and Exergy
By using artificial neural network, exergy eould be predieted through the
eommunity data. The baekpropagation algorithm (Rumelhart et al. 1986) was used
for patterning the input (eommunity) (Fig. 8.22a) and output (exergy) (Fig. 8.22b)
data in a supervised manner. Benthic macroinvertebrate eommunities, eolleeted in
the streams monthly from Suyong River, in Korea, from Oetober 1997 to
September 1998, were used for field data (Fig. 8.22a). Communityeompositions
refleeting the water quality in the Suyong River were reported in Kwon and Chon
(1993), Kang et al. (1995) and Youn and Chon (1999). Fig. 8.22b shows monthly
ehanges in the total exergy for eaeh site in the Suyong and Soktae streams. The
patterns of ehanges in exergy during the survey period oecurred differently
aecording to the sampie loeations. Although the sampie sites were loeated in one
river system and close to eaeh other, the ehanges in exergy showed different
patterns aecording to the sampie sites' loeation and the level of pollution.
Detailed diseussion eould the referred to Park et al. (2001b). Among data for
eommunity and exergy, about one third of the sampies were set aside for testing,
and the rest were used for training.
Training proeeeds on an iterative gradient algorithm, and was similarly eondueted
based on the baekpropagation algorithm as shown previously in the seetion of " •.
A. Multiplayer Pereeptron with Time-Delay". The number of nodes at the hidden
layer used for this study was 5. The learning coeffieient, which updates the
weights at eaeh iteration, was set to 0.7 in this study. The moment eoefficient was
set to 0.8 and the aetivation function eoeffieients were varied between 0.1 and 1.0.
The eonvergenee was generally reaehed in the iteration of 10,000 - 20,000 under
the mean error term of 0.001. Trained data sets were aceordingly matehed to the
original input data in the Soktae Stream (Fig. 8.23a) and in the Suyong Stream
(Fig.8.23b).
T.-S. Chon . Y.S. Park· I.-S. Kwak . E.Y. Cha
et al. 2001b). Data for benthic maeroinvertebrate eommunities (Fig. 8.22a) were
provided for ealeulating exergy, and, aeeording to JjIlrgensen (1997), values
ranging from 29.6 - 43.9 were assigned as W; for maeroinvertebrates. In available
information sourees (JjIlrgensen 1997; Fonseca et al. 2000), however, only values
of W; at higher taxa are available, and the values for the order or family level in
benthic maeroinvertebrates are not specifically provided. Based on experienees
from data analyses and field experiences we arbitrarily assigned 30 for
Oligoehaeta, Diptera, Chironomidae and Hirudinae, and 35 for Gastropoda,
Ephemeroptera, Pleeoptera, Trichoptera, Odonata, and Megoloptera in this study
(Park et al., 2001b).
Training between Community and Exergy
By using artificial neural network, exergy eould be predieted through the
eommunity data. The baekpropagation algorithm (Rumelhart et al. 1986) was used
for patterning the input (eommunity) (Fig. 8.22a) and output (exergy) (Fig. 8.22b)
data in a supervised manner. Benthic macroinvertebrate eommunities, eolleeted in
the streams monthly from Suyong River, in Korea, from Oetober 1997 to
September 1998, were used for field data (Fig. 8.22a). Communityeompositions
refleeting the water quality in the Suyong River were reported in Kwon and Chon
(1993), Kang et al. (1995) and Youn and Chon (1999). Fig. 8.22b shows monthly
ehanges in the total exergy for eaeh site in the Suyong and Soktae streams. The
patterns of ehanges in exergy during the survey period oecurred differently
aecording to the sampie loeations. Although the sampie sites were loeated in one
river system and close to eaeh other, the ehanges in exergy showed different
patterns aecording to the sampie sites' loeation and the level of pollution.
Detailed diseussion eould the referred to Park et al. (2001b). Among data for
eommunity and exergy, about one third of the sampies were set aside for testing,
and the rest were used for training.
Training proeeeds on an iterative gradient algorithm, and was similarly eondueted
based on the baekpropagation algorithm as shown previously in the seetion of " •.
A. Multiplayer Pereeptron with Time-Delay". The number of nodes at the hidden
layer used for this study was 5. The learning coeffieient, which updates the
weights at eaeh iteration, was set to 0.7 in this study. The moment eoefficient was
set to 0.8 and the aetivation function eoeffieients were varied between 0.1 and 1.0.
The eonvergenee was generally reaehed in the iteration of 10,000 - 20,000 under
the mean error term of 0.001. Trained data sets were aceordingly matehed to the
original input data in the Soktae Stream (Fig. 8.23a) and in the Suyong Stream
(Fig.8.23b).