Evaluation of discrete choice and neural network approaches for modelling driver compliance with traffic information

Abstract

This article evaluates dynamic driver behaviour models that can be used, in the context of intelligent transport systems (ITS), to predict drivers’ compliance with traffic information. The inputs to this type of models comprise drivers’ individual socio-economic characteristics and other variables that may influence their compliance behaviour. The output is a binary integer representing whether drivers comply with travel advice or not. Two approaches are available for formulating this category of classification problems: discrete choice models and artificial neural networks (ANNs). The literature on this topic clearly points to the limitations of the discrete choice approach which suffers from assumptions of perfect information about travel conditions, infinite information processing capabilities of drivers and inability to model the uncertainty in driver decision making or the vagueness in information received from ITS devices. ANNs, on the other hand, are able to deal with complex non-linear relationships, are fault tolerant in producing acceptable results under imperfect inputs and are suitable for modelling reactive behaviour which is often described using rules, linking a perceived situation with appropriate action. This study aims to evaluate the performance of these two categories of models based on a common data set of driver behaviour, collected from a field behavioural survey on a congested commuting corridor in Brisbane, Australia. This article proposes the combination of fuzzy logic and neural networks as a viable approach for overcoming the limitations of existing algorithms by modelling drivers as heterogeneous individuals. The results showed superior performance for a neuro-fuzzy model over binary choice models in terms of classifying or predicting the categories of drivers most likely to comply (or not comply) with traffic advice. The accuracy of the proposed model, in terms of classification rate, ranged between 95 and 97% compared to 50–73% for the discrete choice models.