Satellite-based ITS Data Offloading & Computation in 6G Networks: A Cooperative Multi-Agent Proximal Policy Optimization DRL with Attention Approach

TL;DR

This paper presents a novel satellite-based data offloading and computation framework for 6G ITS, utilizing DRL with attention to optimize resource allocation and improve performance over traditional methods.

Contribution

It introduces a cooperative multi-agent DRL approach with attention for satellite-based ITS data offloading, addressing a complex NP-hard optimization problem.

Findings

Performance improvements of up to 9.9% over baselines

Effective resource allocation via DRL with attention

Validated through extensive simulations

Abstract

The proliferation of intelligent transportation systems (ITS) has led to increasing demand for diverse network applications. However, conventional terrestrial access networks (TANs) are inadequate in accommodating various applications for remote ITS nodes, i.e., airplanes and ships. In contrast, satellite access networks (SANs) offer supplementary support for TANs, in terms of coverage flexibility and availability. In this study, we propose a novel approach to ITS data offloading and computation services based on SANs. We use low-Earth orbit (LEO) and cube satellites (CubeSats) as independent mobile edge computing (MEC) servers that schedule the processing of data generated by ITS nodes. To optimize offloading task selection, computing, and bandwidth resource allocation for different satellite servers, we formulate a joint delay and rental price minimization problem that is mixed-integer non-linear programming (MINLP) and NP-hard. We propose a cooperative multi-agent proximal policy optimization (Co-MAPPO) deep reinforcement learning (DRL) approach with an attention mechanism to deal with intelligent offloading decisions. We also decompose the remaining subproblem into three independent subproblems for resource allocation and use convex optimization techniques to obtain their optimal closed-form analytical solutions. We conduct extensive simulations and compare our proposed approach to baselines, resulting in performance improvements of 9.9%, 5.2%, and 4.2%, respectively.