A Machine-learning approach to queue length estimation using tagged customers emission

Dmitry Efrosinin, Vladimir Vishnevsky, Natalia Stepanova
In this paper, we consider the problem of the queue length estimation if only some small number of a so-called tagged customers is observable. The problem is treated in terms of the queueing of vehicles behind a traffic light. A supervised machine learning, particularly an artificial neural network, is used to construct non-linear relationships between the feature and the target. For data generation we simulate an appropriate queueing system. We used an auxiliary Fourier series correction factor by training the neural network. As a result, the quality of the queue length estimation expressed in form of the empirical distribution function of an absolute error was considerably improved.