Information Technology Reference
In-Depth Information
Figure 9. Supply chain demand modeling neural network design
Supply Chain Demand Modeling Neural Network Design
w
1
Neuron
1
Input
1
w
2
w
1
Bias
Neuron
2
Input
2
w
2
Neuron
1
Future
Demand
Bias
w
3
Neuron
3
Input
3
Bias
Bias
w
(h*o)+o
Legend:
h = Hidden Layer Neurons
i = Inputs
o = Outputs
w = Weights
Neuron
h
Input
i
w
(n*h)+h
Bias
and avoid oscillations in the learning space. In addition to the variable learning rate, our first neural
network learning algorithm also included the momentum (Hagan et al., 1996).
To help the neural network stop training before it overfits the training set to the detriment of gen-
eralization, we use a cross-validation set for early stopping. This cross-validation set is an attempt to
estimate the neural network's generalization performance. As previously presented, based on the amount
of data available, we have defined the cross-validation (CV) set as the last 20% of the training set. This
set is removed from the training set and is verified after every training epoch. An epoch is a single cycle
over all of the training data. The error on the cross-validation set will decrease as the network starts to
learn general patterns and then will increase as the network starts to memorize the training set. Thus,
the weights that resulted in the lowest error rate on the cross-validation set are identified as the neural
network model that provides the best generalization performance.
An example graph of the training and cross-validation set errors is presented in Figure 10 where
we see the training set error as a dark shaded line and the cross-validation set error as a light shaded
line. The y-axis represents the error and the x-axis represents the epochs, so Figure 10 presents a visual
representation of the error minimization as the neural network learns through the epochs. The example
Search WWH ::
Custom Search