Bayesian Trained Neural Networks to Forecast Travel Times

This paper will discuss how short-term forecasting of travel time is one of the central topics in current intelligent transportation systems (ITS) research and practice. The most widely applied travel time forecasting approach is the neural network. Usually many candidate neural networks care trained and the network performing best on an independent dataset is selected. However with this approach the training data needs to be divided in two, leading to less well trained neural networks. Using Bayesian inference theory, an ‘evidence’ factor, which can be used to select high-performance networks, is calculated for each network, without the need for a validation set. All training data can then be used for learning of the weights, which results in higher prediction accuracy. A case of forecasting travel times on the A12 motorway in the Netherlands shows that both methods may be preferred in different situations. The evidence approach can be favored over the validation set approach if data are scarce as more data can then be used to train the network. If a wealth of data is available, the validation set approach may be favored.


  • English

Media Info

  • Media Type: Print
  • Features: Figures; References; Tables;
  • Pagination: pp 79-94
  • Monograph Title: TRAIL in Perspective. Proceedings 2008, 10th International TRAIL Congress

Subject/Index Terms

Filing Info

  • Accession Number: 01140608
  • Record Type: Publication
  • ISBN: 9789055841127
  • Files: TRIS
  • Created Date: Sep 21 2009 8:47PM