Forecasting Supply Chain Demand Using Machine Learning Algorithms - Distributed Artificial Intelligence, Agent Technology, and Collaborative Applications - page 346

(Elman, 1990). The recurrent connections feed information from the past execution cycle back into
the network. This permits a neural network to learn patterns through time. Thus the same recurrent
networks with the same weights and given the same inputs may result in different outputs depending
on the feedback signals currently held in the network. In our experiments we will use the two previ-
ously described training methods, the variable learning rate with momentum and early stopping based
on cross-validation set error and the Levenberg-Marquardt with Automated Bayesian Regularization
training algorithm. The addition of recurrent connections also increases the size of the network by the
number of hidden layer neurons squared.

Figure 12. Recurrent subset of neural network design

Recurrent Subset of Supply Chain Demand Modeling
Neural Network Design

rw 1 Neuron 1

rw 2

Bias

Neuron 2

Legend:
h = Hidden Layer Neurons
rw = Recurrent Weights

Bias

Neuron 3

Bias

Neuron h

rw (h*h)

Bias