Identifiability and Practical Relevance of Complex Car-Following Models
This article looks at car-following models from a deliberately pragmatic perspective: What information about driver behavior can be extracted from a given data set without more or less speculative assumptions about underlying behavioral laws? The objectiv
- PDF / 312,198 Bytes
- 9 Pages / 439.36 x 666.15 pts Page_size
- 104 Downloads / 193 Views
stract This article looks at car-following models from a deliberately pragmatic perspective: What information about driver behavior can be extracted from a given data set without more or less speculative assumptions about underlying behavioral laws? The objective of this exercise is not to invalidate existing models but to obtain a better understanding of how much (complex) model structure can be revealed/validated from real data.
1 Introduction The estimation of parameters of a microscopic traffic flow model appears at first glance to be a technically straightforward and well understood procedure. What is not that well understood is the question what is actually revealed by the calibration. Typically, the calibration exercise results in parameters that minimize some distance between model outputs and reality. Some of these parameters have immediate physical meanings: maximum speed, maximum acceleration, and the like. Other parameters are hard to interpret and hence are difficult to validate in hindsight, not even through simple plausibility checks. While sophisticated car-following model specifications abound, much of their added value lies in theoretically being able to explain certain (rare) phenomena, which, however, are of limited relevance if one is interested in estimating, say, a car-following model component for a complex network simulation from real data.
G. Flötteröd () KTH Royal Institute of Technology, Stockholm, Sweden e-mail: [email protected] P. Wagner Y.-P. Flötteröd DLR German Aerospace Center, Cologne, Germany e-mail: [email protected]; [email protected] V.V. Kozlov et al. (eds.), Traffic and Granular Flow ’11, DOI 10.1007/978-3-642-39669-4__5, © Springer-Verlag Berlin Heidelberg 2013
47
48
G. Flötteröd et al.
This article adopts an (almost naive) engineering perspective on the problem in that it starts the analysis by estimating a set of simple linear models from a given data set. If a linear model already explains the data well, there is little reason to complexify the model further. If, however, a linear model fails to explain certain aspects of the data, it still is possible to analyze the residuals in order to obtain datadriven hints of how to improve the model. Clearly, an argument against this approach is that a good model structure, based on physical and/or behavioral considerations, should also have a superior explanatory power. The counter-argument has in essence already been phrased: A simple model with only few, interpretable parameters may result in a slightly inferior fit, but its estimation may be more robust than for a complex model, and its simplicity and interpretability is likely to be a key feature in its practical application. The remainder of this article investigates the above claims by constraining itself to utmost simple models and exclusively inferring model structure either from exact physical laws or from the data itself. Section 2 estimates and analyzes a set of linear car following models. Section 3 then discusses the implications of t
Data Loading...
