Multilayer Artificial Neural Networks for Level-of-Service Analysis of Signalized Intersections

Abstract

The effects of architecture, learning mode, and learning rate on the performance of a level-of-service (LOS) analysis model using an artificial neural network (ANN) are discussed. Multilayer LOSANN models demonstrated improved quality of learning and testing over single-layered models in evaluating level of service of signalized intersections given geometric, traffic, and traffic signal control data. At present, LOSANN takes delay data from Highway Capacity Software (HCS) outputs; hence its accuracy is constrained by the accuracy of the HCS analyses. However, if delays can be determined directly by field observation, the relationships (or patterns) between field-measured delays and the traffic, geometric, and signal control conditions can be fed to LOSANN. Then the neural network-based model can evaluate the level of service at a higher level of accuracy, and such models can be used as part of advanced traffic management systems to automate LOS analyses.