Self-Attention Enhanced Multi-Agent Reinforcement Learning for UAV Trajectory Design in MEC Networks

Abstract

Nowadays, Unmanned Aerial Vehicle (UAV)-assisted Mobile Edge Computing (MEC) has been widely recognized as a promising technique for flexibly and effectively handling computation-intensive and delay-sensitive tasks in 5G advanced and 6G networks. This paper investigates the joint optimization of UAV trajectory and computation offloading for postdisaster rescue scenarios. Specifically, we aim to minimize the task processing delay by jointly optimizing UAV flight trajectory and task offloading decisions. To deal with the problem, a self-attention based Multi-Agent Proximal Policy Optimization (MAPPO) algorithm is designed to optimize the multi-UAV trajectories. Under the Centralized Training with Decentralized Execution (CTDE) framework, this algorithm utilizes a multi-head self-attention mechanism. Each UAV is regarded as an agent, and a self-attention encoder is innovatively integrated at the front-end of each Actor network. The relative importance between each agent and all users is dynamically captured, and the context-aware features are extracted to make more intelligent and collaborative trajectory designs. Through extensive experiments, the proposed algorithm achieves a 16% reduction in total processing latency with respect to the proximal policy optimization policy.