QUAD: QUantum Assisted Distributed Deep Reinforcement Learning for Offloading in NTN-empowered MEC

Abstract

In this paper, we propose QUAD, a quantum assisted distributed deep reinforcement learning (DRL) for offloading in Non-Terrestrial Networks (NTN) empowered multi-access edge computing (MEC) systems. QUAD aims to improve task offloading efficiency by leveraging a quantum layer with hybrid DRL to handle latency-sensitive tasks amid dynamic resource constraints. QUAD utilizes multi-agent distributed deep reinforcement learning (MADDRL) framework based on centralized training with decentralized execution (CTDE), augmented by quantum layers that provide superior global state representations. This enables entanglement-correlated value estimation across agents while maintaining lightweight decentralized execution for offloading decisions. QUAD employs variational quantum circuits (VQC) to optimize neural network parameters to explore high-dimensional state-action spaces. Simulation results demonstrate a 46% reduction in cost calculated as the weighted sum of delay and energy consumption, and a decrease in the task drop rate from 10% to 0% compared to the classical offloading baseline schemes, and a smoother, more stable policy with faster convergence, validating the efficacy of the proposed framework.