Vehicle Edge Computing Task Offloading Strategy Based on Multi-Agent Deep Reinforcement Learning

Vehicle Edge Computing Task Offloading Strategy Based on Multi-Agent Deep Reinforcement Learning

연구 분야: Networking

학회: Journal of Grid Computing

With the rapid advancement of autonomous driving technology, the volume of data generated by vehicle sensors is growing exponentially. Due to spatial limitations within vehicles, high-performance computing equipment cannot be accommodated, resulting in limited vehicle resources being unable to meet the real-time and low-energy consumption demands of vehicular network data processing. To address this challenge, Vehicular Edge Computing (VEC) has emerged as a new development trend in intelligent transportation. VEC effectively relieves the pressure of data transmission and processing by integrating edge computing into the in-vehicle network. Based on this, an efficient offloading strategy for VEC system under resource-constrained conditions is proposed. Firstly, a vehicular edge computing system framework incorporating Non-Orthogonal Multiple Access (NOMA) and queuing theory is constructed, involving multiple servers and multi-priority tasks. The transmission model and computation model from vehicles to edge nodes are derived. This framework significantly reduces inter-vehicle signal interference using the characteristics of NOMA, and optimizes the average delay and energy consumption performance during multi-task processing. Secondly, considering the resource constraints in practical operations, this paper establishes a stochastic game with delay, energy consumption, and cost as the optimization objectives. Given the superiority of Multi-Agent Deep Reinforcement Learning (MADRL) in complex environments, a MADRL-based task offloading algorithm, named Multi-Agent Distributed Deep Action Semantic Deterministic Policy Gradient (MADSASPG), is proposed for vehicular task computation offloading. Finally, to further enhance the action prediction accuracy and performance of the algorithm, the Actor network of the algorithm is improved into an Action Semantic Network (ASN). Through information sharing among Agents, the cooperation between Agents is strengthened. The experimental simulation results show that MADSASPG has good convergence, scalability, and robustness compared to several other algorithms in a multi-intelligent body environment with multiple edge nodes and multiple tasks, as well as good performance in terms of task latency, energy consumption, cost, and server utilization.