RoamML: Distributed Machine Learning at the Tactical Edge

Simon Dahdal, Filippo Poltronieri, Alessandro Gilli, Mauro Tortonesi, Roberto Fronteddu, Raffaele Galliera, Niranjan Suri

Abstract: Machine learning is an effective methodology for enabling real-time data-driven decision-making in tactical scenarios, but its application in such scenarios raises many challenges due to data volume, unpredictable connectivity, and infrastructural challenges in edge environments. Furthermore, the need to perform training operations on remote powerful computing nodes might not be suited for tactical edge networks that often lack high bandwidth links, thus causing critical delays in assessing relevant information for decision-making. To overcome these challenges and enable machine learning at the tactical edge, this paper presents RoamML, a novel distributed continual learning approach tailored explicitly for the tactical edge. Built upon the foundational principle that "moving the model is usually much cheaper than transferring extensive datasets", RoamML seamlessly performs training operations by traversing network nodes, according to the data gravity concept. As RoamML encounters new data at each node, it continually trains on the encountered data to ensure that RoamML maintains an up-to-date and accurate model without transferring data to a centralized entity. Experimental results comparing RoamML with a baseline centralized machine learning solution show the potential of the proposed approach, which is capable of closely matching the accuracy of the baseline method.