Modelling contamination sites with trainable networks Conference

cited authors

  • Tansel, I; Bao, WY; Tansel, B; Jordahl, CM

abstract

  • The modelling performance of Backpropagation-type neural networks (BP-NN) and Abductory Induction Mechanism (AIM) were studied. Both trainable networks were used to model a mathematical function and the detectable levels of methyl tertiary butyl ether (MTBE) data of South Miami. Available sample size of the MTBE data was very small and required multiple modelling. For n samples, the modelling was repeated n times by using both trainable networks. n training sets were prepared with n-1 cases. One of the eight cases was missing in each of these training sets. BP-NNs and AIM were used to model the data. For each approach, three different models were prepared by selecting the number of hidden nodes or penalty functions. The developed models were tested on the nth case network has never seen before. In the study, 24 BP-NN and 24 AIM models were tested. The AIM was found convenient and fast for modelling. The BP-NN took a long time for training and required the input of the operator to test different hidden layer structures. The accuracies of both methods were very close to each other.

publication date

  • December 1, 1995

start page

  • 823

end page

  • 828

volume

  • 5