Description
Date depot: 10 février 2021
Titre: Uncertainty and data quality in simulations
Directeur de thèse:
Paolo PAPOTTI (Eurecom)
Encadrant :
Motonobu KANAGAWA (Eurecom)
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Non defini
Resumé: Several systems use AI and algorithms to aid traffic engineers in optimizing traffic. The challenge of the adoption of AI is that it explores a wider set of technical options and parameters w.r.t. traditional traffic optimization systems, and thus requires reliable validation on simulated scenarios to allow risk management and containment. Traffic simulation tools are a type of technology developed for simple and human-supervised proof-of-concept to validate intuitions and ideas of traffic engineers responsible to develop and evolve traffic infrastructure in small part of a city. Simulation results are part of a more comprehensive set of evaluations parameters that help decisions makers in driving urban and road infrastructure evolution. However, in the case of AI validation, all the limitations of simulation tools need to be unfolded and measured as they are one of the source of uncertainty in the validation process. Moreover, AI heavily depends on data, but data about traffic is the result of an imperfect detection process running at scale. In most settings, traffic data are the result of a continuous video processing aimed at identifying moving objects (e.g. vehicle, bikes, pedestrian) and detect their trajectories. However, the AI-based video analysis is also subject to error (false positive and negative detections, as well as mis-classification of the objects). The same noisy data is also used to configure and calibrate traffic generation into a simulated scenario, providing another source of uncertainty related to the validation of AI algorithms controlling traffic optimization in a simulated environment.
Doctorant.e: Mitsuzawa Kensuke