An MARL-based Task Scheduling Algorithm for Cooperative Computation in Multi-UAV-Assisted MEC Systems

This paper studies the task scheduling problem in a multi-UAV-assisted system and formulates the problem as an optimization problem with an objective to maximize the number of tasks successfully completed in the system. To solve the formulated problem, we model the optimization problem as a decentralized partially observable Markov decision process (Dec-POMDP) and propose an MARL-based task scheduling (MTS) algorithm to optimize scheduling decisions for computing tasks arriving in the system. The MTS algorithm operates in a centralized training and decentralized execution manner based on QMIX. Using the MTS algorithm, each UAV makes scheduling decisions in accordance with its local observation and local action policy in a decentralized manner while the local action policies of all UAVs are jointly trained by a central controller in a centralized manner. Moreover, the MTS algorithm leverages recurrent neural networks (RNNs) to deal with the partial observability of a POMDP. Simulation results show that the proposed MTS algorithm can converge faster and improve the number of tasks successfully completed in the system compared with benchmark algorithms.