This paper summarises the findings of an initial study in which neural networks were developed for several kinds of transport problem. The first problem addressed was that of recognising whether the road system is in a particular state (e.g. that certain links are "congested"). The second problem was similar, but concerned with short term forecasting from real time data; some comparisons with Box-Jenkins methods are reported. Because certain traffic states or parameters (e.g. queue lengths) are not readily measurable on-street, an experiment is also reported where a neural network is used to infer relationships between these parameters and more readily measurable quantities. Data for this experiment were provided by micro-simulation. Such modelling capabilities were investigated further in the fourth area, by applying neural networks to complex multivariate data (here a computer-based survey of drivers' route choice); it was found that neural networks provided an effective and quicker alternative to logit models of individual choice. However, there remained the worry that the relationships derived are not made explicit by the neural network. This shortcoming was circumvented in the fifth study, in which it was shown that by dithering the data inputs, useful insights were obtained into the underlying structure of the relationships. (A)
Abstract