Methodological issues in building, training, and testing artificial neural networks in ecological applications

Creative Commons License

Ozesmi S. L., Tan C. O., Ozesmi U.

ECOLOGICAL MODELLING, vol.195, pp.83-93, 2006 (SCI-Expanded) identifier identifier

  • Publication Type: Article / Article
  • Volume: 195
  • Publication Date: 2006
  • Doi Number: 10.1016/j.ecolmodel.2005.11.012
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus
  • Page Numbers: pp.83-93
  • Keywords: artificial neural networks, back-propagation, modelling, model, model generalizability, ecology, MODELS, REGRESSION, PREDICTION, ABUNDANCE, FISH
  • Middle East Technical University Affiliated: No


We evaluate the use of artificial neural networks, particularly the feedforward multilayer perceptron with back-propagation for training (MLP), in ecological modelling and make suggestions on its use. in MLP modelling, there are no assumptions about the underlying form of the data that must be met as in standard statistical techniques. Instead, researchers must clarify the process of modelling, as this is most critical to how the model performs and is interpreted. Overfitting on the data, a potential problem, can be avoided by limiting the complexity of the model and by using techniques such as weight decay, training with noise, and limiting the training of the network. Methods on when to stop training include: (1) early stopping based on cross-validation, (2) stopping after a analyst defined error is reached or after the error levels off, and (3) use of a test data set. The third method is not ideal as the test data set is then not independent of model development and the resulting model may have little generalizability. The importance of an independent data set cannot be overemphasized as we found dramatic differences in model accuracy assessed with prediction accuracy on the training data set, as estimated with bootstrapping, and from use of an independent data set. The comparison of the artificial neural network with a general linear model (GLM) as a standard procedure is recommended because a GLM may perform as well or better than the MLP. In such cases, there are no interactions or non-linear terms that need to be modelled and it will save time to use the GLM. Techniques such as sensitivity analyses, input variable relevances, neural interpretation diagrams, randomization tests, and partial derivatives should be used to make MLP models more transparent, and further our ecological understanding, an important goal of the modelling process. Based on our experience we discuss how to build an MLP model and how to optimize the parameters and architecture. The process should be explained explicitly to make the MLP models more readily accepted by the ecological research community at large, as well as to make it possible to replicate the research. (c) 2005 Published by Elsevier B.V.