Abstract
Vehicular Edge Computing (VEC) integrates the computing resources of vehicles to provide services for other vehicles and pedestrians through task offloading. However, the vehicular environment is characterized by high mobility, uncertain wireless channels, and dynamic workloads, which complicate offloading decisions. This paper proposes the Adaptive Learning-based Task Offloading (ALTO) algorithm, a distributed solution based on Multi-Armed Bandit (MAB) theory. ALTO enables vehicles to learn the delay performance of neighbors without frequent state exchanges. By incorporating input-awareness and occurrence-awareness, ALTO adapts to varying task sizes and changing network topologies. Simulations demonstrate that ALTO achieves sublinear learning regret and reduces average offloading delay by up to 30% compared to existing learning-based methods.
Introduction
Vehicular Edge Computing (VEC) aims to transform vehicles into mobile data centers. In this paradigm, a "Task Vehicle" (TaV) with heavy computational requirements—such as autonomous driving sensor fusion—can offload tasks to nearby "Service Vehicles" (SeVs) with available CPU cycles.
Despite its potential, VEC faces significant challenges. High mobility causes rapid changes in network topology, and the uncertainty of wireless channels and CPU loads makes it difficult for a TaV to select the optimal SeV. Traditional optimization often requires massive signaling overhead to maintain state awareness. This work addresses these challenges by framing task offloading as a learning problem where vehicles optimize their decisions through experience.
The ALTO Algorithm
The core of the proposed solution is the ALTO algorithm, which treats neighboring SeVs as "arms" in a Multi-Armed Bandit (MAB) framework. In a standard MAB, an agent balances exploring new options with exploiting known high-reward options. ALTO enhances this with two key innovations:
- Input-Awareness: ALTO recognizes that the cost of a sub-optimal decision is proportional to the task size. It uses smaller tasks as low-risk opportunities to explore less-known SeVs, while prioritizing exploitation of the best-known providers for large, delay-sensitive tasks.
- Occurrence-Awareness: Unlike static MAB models, the set of available SeVs in VEC is constantly changing. ALTO tracks the "birth time" of each SeV, ensuring that newly arrived vehicles are given a fair chance to be evaluated against established ones.
Because ALTO is distributed, each TaV maintains its own local estimates of SeV performance based on observed total delays (upload, execution, and download), eliminating the need for constant, bandwidth-heavy signaling.
Performance Evaluation
The authors evaluated ALTO using both synthetic simulations and a realistic highway scenario. The realistic simulation utilized the Veins framework, coupling SUMO for traffic mobility and OMNeT++ for wireless modeling. They modeled a 12km segment of the G6 Highway in Beijing using OpenStreetMap data.
Key Results
- Delay Reduction: In highway simulations, ALTO reduced average offloading delay by up to 30% compared to the standard Upper Confidence Bound (UCB) algorithm.
- Learning Efficiency: ALTO outperformed existing variants like VUCB and AdaUCB by 30% to 65% in terms of cumulative learning regret. The algorithm achieves "sublinear regret," meaning its performance converges toward the theoretical optimum over time.
- Fast Convergence: As shown in the performance plots, ALTO reaches a stable, near-optimal state significantly faster than competing algorithms, a critical requirement for highly mobile vehicular environments.
Conclusion
ALTO provides a robust, distributed mechanism for task offloading in VEC systems. By shifting the focus from frequent signaling to adaptive learning, it effectively manages the inherent uncertainties of the vehicular environment. Future work may explore joint optimization of vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) offloading, as well as scenarios involving multiple competing Task Vehicles.
Summary updated