Differentiable Optimization for Orchestration: Resource Offloading for Vehicles in Smart Cities

• Published in: IEEE access Vol. 12; p. 1
• Main Authors: Strauss, Thilo, Oechsle, Michael, Bauknecht, Uwe
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.01.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Automobiles; Cities; Cloud computing; Combinatorial analysis; Connected and Autonomous Vehicle (CAV); Constraints; Edge computing; Heuristic methods; Indexes; Infrastructure; Intelligent sensors; Mobile computing; Multi-access Edge Cloud (MEC); Optimization; Orchestration; Roads; Roadsides; Smart cities; Smart City; Task analysis; Task complexity; Wireless communication
• DOI: 10.1109/ACCESS.2024.3363426
• ISSN: 2169-3536

Connected and Autonomous Vehicles (CAV) which interact with Roadside Units (RSU) as part of a smart city infrastructure are currently seeing first real-world deployments. Not only can CAVs benefit from access to a cities' infrastructure by obtaining data from various sensors (e.g., Video or Lidar), but they can also leverage the broad network coverage to offload complex computation tasks from their limited on-board hardware to scalable cloud resources. Furthermore, a smart city supporting multi-access edge computing (MEC) can even provide safety-relevant and time-critical services thanks to reduced latency and increased reliability. This requires an algorithm to determine which vehicle offloads computation to which computation resource in the city. This orchestration task is a challenging combinatorial problem subject to resource and quality of service constraints. We present a novel and powerful, yet surprisingly simple algorithm that provides a good and fast approximation to this problem. This Differentiable Orchestrator converts a combinatorial problem into a soft-constrained differentiable analog, which can be solved very quickly. We compare the proposed method with other heuristic methods and conclude that it significantly outperforms most competing methods in artificial examples and realistic scenarios. In order to make the method as reproducible as possible and serve as a baseline for future research we make our data and simulations freely available.