Obtain an Optimum Artificial Neural Network Model for Reservoir Studies

Abstract Artificial Neural Networks (ANNs) excel in dealing with uncertainty, fuzziness, incompleteness, and poorly defined nonlinear systems. These factors widely exist in reservoir studies. Training neural networks is a notoriously difficult problem. In training neural networks, one of the major pitfalls is overtraining, analogous to curve fitting for rule-based systems. Emperical evidence suggests that the number of records must exceed the number of neural network weights by a factor of two to minimize overtraining problems. Undertraining is another problem in which the ANNs with a simpler architecture cannot master the basic rules of input patterns. The aftermath of overtraining is much worse than that of undertraining. Based on the trial and error method, this paper first explores the overtraining problem using various defined functions and then applies the results to an oil well in the Nash Draw field in New Mexico. An optimum architecture was found for the field problem. Solutions to minimize neural network overtraining problems are presented.