Page 277 - PVT Property Correlations

Artificial Neural Network Models for PVT Properties Chapter | 10 243

Optimization Algorithm
The learning process of the ANN requires thousands or sometimes millions
of error minimization iterations. An optimization routine is used to find the
weights of the ANN. The choice of the optimization algorithm depends on
the nature of the objective function and the engineering problem or applica-
tion. Examples of common optimization algorithms used in ANN include
gradient directive methods, nonlinear least squared, genetic algorithms, like-
lihood statistical methods, and many others.

Training Control Parameters

During the training process, some parameters are used to minimize the train-
ing time as much as possible, with the same level of required error quality.
The parameters are learning rate, weight decay factors, and node momentum.
Learning rate is used to control the maximum change allowed for any of the
node weights during the training process. The decay factor is used to control
the minimum change of the weights. The momentum of the node is used to
signify or designify the node value after or before activation.

Running Control Parameters

Running control parameters are parameters that can be used to terminate the
training process. Running control parameters include maximum number of
iteration, minimum required network error, and minimum tolerance in weight
value changes. Running control parameters affect the time required for train-
ing and ANN prediction accuracy.

FUTURE DEVELOPMENT AND EXPECTATIONS
Although many ANN models have been developed to predict different PVT
properties over the past several years, additional models are expected to
appear in the literature. Other competing artificial intelligence (AI) and
machine learning techniques (e.g., relevance vector machines, support vector
machines, and random forests) are anticipated to receive growing attention in
prediction of PVT properties. Like the correlation models developed by
regression techniques, ANN models and others will be applicable within the
ranges of the data that are used to develop them.
Improved models are expected to appear when more PVT data records
are incorporated in the training of these models. The data must be quality
controlled, and relations between input and output values must exist. The
strength of the ANN models lies in their ability to find the relations between
inputs and outputs. These relations may not be obvious or clear (i.e., not
showing a trend).