Trip generation: comparison of neural networks and regression models

Modelling the number of trips produced by the inhabitants of a zone, the trip generation, is complex and highly dependent on the quality and availability of data. It seems almost impossible to model/forecast the number of trips a person makes without adequate amounts of data. Transportation engineers are commonly faced with a question that is related to this topic; how to perform reliable trip generation with scarce and expensive field data. It is therefore interesting to find the method that gives the best results with the smallest data sets. This paper deals with trip generation and explores the performance of neural networks and commonly used regression models. This research tries to answer the question whether neural networks can out-perform traditional regression methods or not. The neural networks are tested in two situations with regards to the data availability; (i) data is scarce; and (ii) data is sufficiently at hand. Synthetic households, generated using travel diary data, are the basis for the research. These households are divided over a zone in varying complexities, from homogeneous without statistical deviation on the household characteristics to inhomogeneous with a deviation on the household characteristics. The use of synthetic data, without unknown noise, gives the opportunity to clearly determine the impact of complexity on the forecasting results. The question of whether neural networks can be used in trip generation modelling is answered positively. However, neural networks do not overall out-perform classical regression models in situations where data is scarce. The advantages over regression models are negligible.