Page 275 - PVT Property Correlations

Artificial Neural Network Models for PVT Properties Chapter | 10 241

Artificial Neural Network Topography
The different ANN topographies and structures are numerous. Some known
network structures perform better in certain applications (e.g., prediction,
optimization, clustering, noise removal, etc.). The most commonly used
ANN structure for prediction problems (including PVT properties prediction)
is the multilayer back propagation structure. The self-organizing map is one
of the most common ANN structures for clustering applications. Clustering
techniques can be used to develop expert systems to automatically select the
best PVT correlations, for example.

Number of Hidden Layers
An ANN can contain any number of hidden layers. The number of hidden
layers can range from zero to a very large number. When multiple hidden
layers are used, the ANN is usually referred to as a deep network. The more
hidden layers, the higher the capabilities of the network to discover the hid-
den relations between inputs and outputs. However, for multilayer networks,
a sufficient number of training records should be available. Training time
will also increase with the number of hidden layers and the number of nodes
in each layer.

Nodes per Layer

The number of nodes in the input and output layers is usually determined by
the number of inputs and outputs in the problem under investigation. The
number of nodes in the hidden layer(s) can be arbitrary. However, the num-
ber of nodes in every hidden layer depends on the ANN application and on
the size of the training data. Use of a different combination of hidden layers
and variation of the number of nodes per layer can be investigated to opti-
mize the network. It is usually recommended, however, to make the number
of nodes per hidden layer a function of the number of nodes in the input
layer. The optimum number of nodes in every hidden layer is a problem-spe-
cific. It is usually determined by trail and error, although some automatic
algorithms have been recently used to find the optimum ANN structure.
Availability of a large number of training records allows the choice of more
nodes in the hidden layer(s) and usually produces networks that are more
accurate.

Layer Connection
One of the ANN parameters that can be used to tune the ANN for a particu-
lar application is the type of layer/node connection. The three distinct types
of connection are (1) full, (2) customized, and (3) recurrent. The type of