TOWARD A ROBUST FRAMEWORK FOR FREEWAY TRAVEL TIME PREDICTION : EXPERIMENTS WITH SIMPLE IMPUTATION AND STATE-SPACE NEURAL NETWORKS

Robustness to missing or faulty input, due to failures in the data collection system, is a key characteristic for any travel time prediction model that is to be applied in a real time environment. Previous research by van Lint et al (2002) has shown that so-called State-Space Neural Networks (SSNN) are capable of accurately predicting experienced travel times. Our paper shows that incorporating corrupt data into the training procedure does increase the robustness of these SSNN models, but at the cost of predictive performance: there is a clear trade off between robustness and model accuracy. More over, inclusion of (small) amounts of corrupted data in the training procedure makes the internal states of the SSNN model which are closely related to the expected traffic conditions, more difficult to interpret. Application of pre-processing (imputation) strategies before feeding the input into the model does lead to a robust but still accurate model. Although there are theoretical shortcomings to simple imputation schemes, the combined framework of pre-processing and SSNN model is robust to various kinds of input failure, even though the proposed pre-processing strategies are naive non-parameterized procedures such as exponential smoothing and spatial interpolation. We hypothesize this is due to the fact that the SSNN model is robust to the "damage" done by the pre-processor. Further research should emphasize on enhancing the proposed pre-processing procedures and applying the travel time prediction framework in real time

• Supplemental Notes:
  • Publication Date: 2003. Transportation Research Board, Washington DC. Remarks: Paper prepared for presentation at the 82nd annual meeting of the Transportation Research Board, Washington, D.C., January 2003. Format: CD ROM
• Corporate Authors:

  University of California, Berkeley

  California PATH Program, Institute of Transportation Studies
  Richmond Field Station, 1357 South 46th Street
  Richmond, CA  United States  94804-4648

  California Department of Transportation

  1120 N Street
  Sacramento, CA  United States  95814

  University of California, Berkeley

  Department of Electrical Engineering and Computer Sciences
  Berkeley, CA  United States  94720
• Authors:
  • Lint, J W C van
  • Hoogendoorn, S P
  • van Zuylen, Henk J
• Conference:
  • Transportation Research Board 82nd Annual Meeting
  • Location: Washington DC, United States
  • Date: 2003-1-12 to 2003-1-16
• Date: 2003

Language

• English

Media Info

• Pagination: 11 p.

Subject/Index Terms

• TRT Terms: Freeways; Neural networks; Traffic estimation; Travel time
• Subject Areas: Data and Information Technology; Highways;

Filing Info

• Accession Number: 00962482
• Record Type: Publication
• Source Agency: UC Berkeley Transportation Library
• Files: PATH, STATEDOT
• Created Date: Sep 2 2003 12:00AM