Multiagent Deep Reinforcement Learning for Decentralized Multi-UAV Mobile Edge Computing Networks

This article studies a new multiagent deep reinforcement learning (MADRL) approach for unmanned aerial vehicle (UAV)-assisted mobile edge computing (MEC) networks, where UAV-mounted servers provide offloading services to mobile users (MUs). We aim to minimize the total energy consumption of MUs by optimizing UAV mobility, UAV-MU association, resource allocation, and task offloading ratios. In the multi-UAV scenario, we model the MEC network as a multiagent partially observable Markov decision process (POMDP), where each UAV agent operates with limited information for decentralized decision-making. Conventional MADRL methods manually design such UAV interaction messages, thereby incurring performance degradation. To address this issue, we propose a new neural network (NN)-based UAV interaction mechanism that generates autonomously task-oriented messages to minimize energy consumption. Such message-generating NNs are developed under the MADRL framework, which allows for joint optimization of UAV interactions and decentralized decisions in an end-to-end manner. Numerical results demonstrate that our approach outperforms traditional MADRL methods and achieves performance close to ideal centralized schemes while maintaining scalability with varying UAV numbers.