Page 329 - Data Science Algorithms in a Week

310 Mayra Bornacelli, Edgar Gutierrez and John Pastrana

Figure 5. Schematic of a neural network.

Selection of an Appropriate Architecture for Neural Networks to Predict
the Price of Thermal Coal

An appropriate architecture for the neural network (i.e., the number of neurons in the
hidden layer) had to be selected since the backpropagation algorithm was used. Moody
and Utans (1992) indicated that the learning ability (CA) of a neural network depends on
the balance between the information of the examples (vectors) and the complexity of the
neural network (i.e., the number of neurons of the hidden layers - which also tells us of
the number of weights since they are proportional). It is important to say that a neural
network with few weights and therefore neurons in the hidden layers (λ) will not have the
proper CA to represent the information of the examples. On the other hand, a neural
network with a large number of weights (i.e., degrees of freedom) will not have the
capability to learn due to overfitting.
Traditionally in supervised neural networks CA is defined as expected performance
in data that is not part of the training examples. Therefore, several architectures (different
hidden neurons) are “trained” and the one with the best CA is selected. This method is
especially effective when there are sufficient data samples (i.e, a very large number).