Chapter 8· Analysis of Stream Macroinvertebrate Communities
151
Detailed process for the backpropagation algorithm could be referred to Rumelhart
et al. (1986) and Zurada (1992).
When communities were given as input to the simple multilayer perceptron
with the time delay between one and five months, the convergence was generally
reached in the iteration of 20,000 - 30,000 under the mean error term of 0.05,
which is the sum of square terms of difference in output and target values divided
by the number of input patterns. The training rate leaming and momentum
coefficients were 0.5 and 0.9, respectively. Trained data sets were accordingly
matched to the original input data. It appeared that convergence was dependent
upon the length of time delay (Chon et al. 2000b). When new data were given to
the trained network for recognition, the network was able to make one-step
predictions for the following community in time (Fig. 8.13). In general it
appeared that the predicted and actual field data were in accord, although some
discrepancies were locally observed. The trained results correspondingly reflected
the expected development of communities under the influence of various degrees
of organic pollution in urbanized streams in a certain time span. However, there
were occasions that the degree of correspondences between the actual and
predicted data was not high. It was difficult to obtain precise matching in
densities between the two data sets. This is understandable that predicting density
in "each" taxa of communities in field conditions are generally not easy.
Correlation coefficients between the predicted and field data were 0.556
(P
8.3.2
Elman Network
The temporal development of input data could be revealed by the recurrent neural
network. The Elman type network (Elman, 1990), one of the most well-known
dynamic models in partially connected recurrent learning, was implemented for
learning community dynamics (Chon et al., 2000b). The architecture of Elman
type recurrent neural network (RNN) is basically similar to the multilayer
perceptron except the composition of the hidden layer (Fig. 8.12b). However,
hidden layer embodies another context layer for implementing recurrence.
Recurrence implies that the state of network depends on current input and its own
internal state on the previous cycle. In this case, the hidden layer has recurrence
and its own internal state is represented through the context of the hidden layer.
The number of nodes at the input and output layers was 5, and 30 neurons were
used for the hidden and context layers.
In the input layer, community data for selected Genera, x,(t-l), were given as
external input. Concurrently, output values from the hidden layer for the previous
cycle are also provided as internal inputs to the hidden layer as C/(t-l). Initially,
some small random numbers are used for the internal inputs. The group of x/(t-l)
and C/(t-l) consist of the total input for the hidden layer, z,(t). The sum of linear
combination of weights and inputs, IP-l), is subsequently adjusted in a nonlinear
function such as C/(t-l) = fi~(t-l». The input process could be summarized as
151
Detailed process for the backpropagation algorithm could be referred to Rumelhart
et al. (1986) and Zurada (1992).
When communities were given as input to the simple multilayer perceptron
with the time delay between one and five months, the convergence was generally
reached in the iteration of 20,000 - 30,000 under the mean error term of 0.05,
which is the sum of square terms of difference in output and target values divided
by the number of input patterns. The training rate leaming and momentum
coefficients were 0.5 and 0.9, respectively. Trained data sets were accordingly
matched to the original input data. It appeared that convergence was dependent
upon the length of time delay (Chon et al. 2000b). When new data were given to
the trained network for recognition, the network was able to make one-step
predictions for the following community in time (Fig. 8.13). In general it
appeared that the predicted and actual field data were in accord, although some
discrepancies were locally observed. The trained results correspondingly reflected
the expected development of communities under the influence of various degrees
of organic pollution in urbanized streams in a certain time span. However, there
were occasions that the degree of correspondences between the actual and
predicted data was not high. It was difficult to obtain precise matching in
densities between the two data sets. This is understandable that predicting density
in "each" taxa of communities in field conditions are generally not easy.
Correlation coefficients between the predicted and field data were 0.556
(P
8.3.2
Elman Network
The temporal development of input data could be revealed by the recurrent neural
network. The Elman type network (Elman, 1990), one of the most well-known
dynamic models in partially connected recurrent learning, was implemented for
learning community dynamics (Chon et al., 2000b). The architecture of Elman
type recurrent neural network (RNN) is basically similar to the multilayer
perceptron except the composition of the hidden layer (Fig. 8.12b). However,
hidden layer embodies another context layer for implementing recurrence.
Recurrence implies that the state of network depends on current input and its own
internal state on the previous cycle. In this case, the hidden layer has recurrence
and its own internal state is represented through the context of the hidden layer.
The number of nodes at the input and output layers was 5, and 30 neurons were
used for the hidden and context layers.
In the input layer, community data for selected Genera, x,(t-l), were given as
external input. Concurrently, output values from the hidden layer for the previous
cycle are also provided as internal inputs to the hidden layer as C/(t-l). Initially,
some small random numbers are used for the internal inputs. The group of x/(t-l)
and C/(t-l) consist of the total input for the hidden layer, z,(t). The sum of linear
combination of weights and inputs, IP-l), is subsequently adjusted in a nonlinear
function such as C/(t-l) = fi~(t-l». The input process could be summarized as