3IA PhD/Postdoc Seminar #16

Introduction to the OTESIA institute
 

PINNs approach for traffic model calibration

Abstract: 
The increasing amount of data generated by smartphones and IoT devices motivated the development of Federated Learning (FL), a framework for on-device collaborative and privacy-preserving training of machine learning models. Heterogeneity is a core challenge in Federated Learning: devices with higher availability are likely to participate more often during the training, thus introducing a bias in the learned model. Our theoretical analysis points out a tradeoff between 1) letting everybody participate leading to a faster convergence towards a surrogate of the objective function; 2) correcting the bias resulting in a slower convergence to the true objective. In our work, we propose an energy-efficient solution to alleviate the bias without neglecting the convergence speed.
 

Resource-aware Federated Learning

Abstract:
At the macroscopic scale, traffic is usually represented as a fluid flowing through the road network. To characterize the evolution of the system, aggregate quantities such as the flow Q = Q(t, x) (it veh/h) or the density rho = rho(t, x) (it veh/km) of vehicles are defined. The challenge is to be able to understand, reproduce and anticipate the evolution of density and flow in space and time based on both this mathematical modeling and traffic data. Existing physical traffic flow models, such as Lighthill-Whitham-Richards (LWR) models, which can only capture real-world traffic dynamics to a limited extent, result in low-quality estimation and require massive data for accurate and generalizable estimation. To solve this problem, we use Physics-Informed Neural Networks (PINNs) techniques which have been shown to date to be an efficient numerical tool that provides solutions to partial differential equations (PDEs), even though, theoretically, they have limited ability to solve problems with continuous solutions. Particularly, we focus on the LWR model with observed loop detector data, using traffic density as the traffic variables. We show the advantages and disadvantages of PINNs technique for solving (with loop detector data) LWR physical traffic flow models, and suggest strategies for dealing with the problems that arise when we use neural networks.

Open discussion on the two contributions