.png){kind=icon}
.png){kind=icon}
Return
Computation offloading for vehicular networks via sparse-attention-based multi-agent reinforcement learning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.adhoc.2025.104006.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
For Mobile Edge Computing (MEC) offloading in large-scale Internet of Vehicles (IoV), existing methods struggle to provide low-latency services. This paper proposes a multi-agent algorithm based on sparse attention weighting, aiming to minimize the expected cost. Specifically, a long short-term memory (LSTM) model is integrated into the actor network to assist VUs in predicting the future states of edge servers (ESs). Additionally, a sparse attention mechanism is employed to compress the joint observation space, reducing computational complexity and enabling adaptation to large-scale VU environments. Furthermore, to address the convergence difficulties arising from a large number of agents, curriculum learning is adopted for phased training. We conduct evaluations in a custom IoV simulation environment,experimental results demonstrate that the proposed Multi-Agent Sparse-Attention Soft Actor-Critic (MASASAC) algorithm outperforms baseline methods in both performance and convergence speed, achieving an improvement of approximately 16.4 % to 37.6 % and further enhancing performance by approximately 10.9 % through curriculum learning.