Machine learning methods for transportation under uncertainty

Abstract

The goal of the thesis is to develop and study effective modeling methods for Transportation under uncertainty scenarios. This is motivated by both the prevalence of uncertainty in Transportation and the widespread use of Transportation models in practice, e.g., for traffic management, planning of mobility services and operation of Public Transport. We approach this goal through Machine Learning, namely, our proposed methods extract patterns from data and leverage them for better modeling. The first uncertainty scenario we deal with concerns road incidents. As each incident induces a unique and uncertain change in the correlation structure of traffic variables, prediction models cannot realistically be prepared in advance for every possible incident. We thus devise QTIP: an online framework that uses distress signals from affected vehicles for ad-hoc simulations and corresponding model adaptation, which we experiment in a case study of a Danish highway. The results suggest that QTIP provides a feasible opportunity for mitigating the long-standing problem of instantaneous model adaptation upon incidents. Next, we deal with modeling of mobility demand when its actual value is uncertain, as it is only partially observed. A common reason for this is the use of data from mobility services, whereby demand is observed through limited service supply and is thus “censored”. First, we propose a novel regression method that combines flexible Gaussian Processes with a censored likelihood, and we experiment it on real-world data from bike-sharing and taxi services. The results emphasize that censorship can and should be considered for more accurate modeling of mobility demand. Second, we propose Censored Quantile Regression Neural Networks as an alternative non-parametric method for modeling mobility demand. We also apply this method to real-world data from bike-sharing and shared Electric Vehicles, and the results show that it can outperform both censorship-unaware Neural Networks and censored parametric models. To the best of our knowledge, no previous works have applied Censored Quantile Regression Neural Networks in the Transportation domain. Finally, we study the impact of travel demand uncertainty on the performance of demand-responsive Transport services. First, we devise a framework for predictive optimization, which estimates marginal distributions of travel demand between Origin-Destination pairs, joins them through a copula, and uses the joint distribution for stochastic route optimization. We experiment the framework in a case study of on-campus autonomous mobility, where demand is observed via WiFi-probed crowd movements, and obtain that the framework can outperform conventional optimization techniques that do not leverage the full predictive distribution. Second, we study the impact of prediction inaccuracy via a more general approach, which does not involve specific prediction models, by sampling prediction errors from various distributions. We apply this approach to a case study of demand-responsive Public Transport in the Copenhagen metropolitan area, through which we quantify the relationship between prediction errors and subsequent performance of dynamic routing. In conclusion, this thesis offers several useful findings for Transportation practice and theory. We find that recent technological advances can alleviate the degradation of data-driven prediction models under road incidents, for which we offer a dedicated framework. We also advise to explicitly model the inherent censorship in Transportation demand, for which we offer two non-parametric alternatives. For dynamic operation of shared mobility services, we demonstrate the benefits of preserving a full uncertainty structure of demand, and we also quantify the relationship between predictive quality and subsequent service performance.