Abstract

The use of Artificial Neural Networks (ANN) was investigated as an alternative method to obtain solutions to the Boussinesq equation. First, 52,280 solutions were obtained with a numerical method, each one corresponding to a unique set of parameter values (the inputs). Each solution consisted of two water table level values (the training outputs) corresponding to 12- and 24-h delays. The input/output combinations were then divided into learning and recall sets. Next, 15 fully connected, feed-forward ANNs, having slightly different architectures, were each trained with the learning set for 50,000 cycles. Their performance was then evaluated with the recall set (i.e., the recall outputs obtained from these inputs were compared to the corresponding training outputs) and the most appropriate ANN selected for further training and testing. This “best” ANN was subjected to 21 different training regimes so as to determine the most appropriate one, the regimes consisting of various quantities of training with the learning set. Performance of the ANN at the various training levels was then again evaluated and the optimal regime determined. Finally, the “optimally trained” ANN was transformed into a C program and compiled and linked into a DOS executable file which was then used to calculate solutions to the Boussinesq equation for all 52,280 cases.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call