Enhancing traffic dynamics-induced machine learning through heterogeneous driving policies

Abstract

Transportation systems demand substantial computational resources to support diverse intelligent applications involving vast numbers of mobile agents at the network edge. Existing approaches, such as mobile edge computing, merely redistribute computational tasks to edge devices, relying on in-vehicle computers as computational units. Here, we investigate an alternative computing approach: harnessing the inherent dynamics of physical vehicles without the need of in-vehicle computer as a complimentary and energy efficient computational resource. We propose a physical reservoir computing framework that can leverage dynamics produced by vehicle fleets on roadways and transform them into vast computational resources for various machine-learning (ML) tasks at the edge of network. The proposed framework projects signal inputs to the lead vehicle speed to obtain a nonlinear speed variation of the following vehicles, which is then used as the system readout feeding to a single-layer neural network for linear regression based on a small amount of training data. Our results show a positive correlation between congestion level and accuracy, with heterogeneous driving policies significantly enhancing both. A heterogeneous fleet of 30 vehicles achieved a maximum accuracy of 0.787, comparable to an echo state network with 300 nodes, demonstrating excellent performance across various ML tasks. The results pave the way for utilizing mobile edge agents as novel physical computational resources for various ML tasks, which is crucial for enabling real-time computation close to the data source, significantly improving computing efficiency and latency for transport applications such as autonomous driving and traffic management.