Task Scheduling in Space-Air-Ground Uniformly Integrated Networks with Ripple Effects

TL;DR

This paper addresses task scheduling in space-air-ground integrated networks considering ripple effects, proposing a novel multi-agent reinforcement learning approach to optimize information freshness and energy use.

Contribution

It models ripple effects as a Markov game and applies a modified MAPPO algorithm to effectively manage high-dimensional, partially observable scheduling problems.

Findings

Significant reduction in age of information (AoI) at users.

Lower energy consumption at access points (APs).

Outperforms benchmark methods in simulations.

Abstract

Space-air-ground uniformly integrated network (SAGUIN), which integrates the satellite, aerial, and terrestrial networks into a unified communication architecture, is a promising candidate technology for the next-generation wireless systems. Transmitting on the same frequency band, higher-layer access points (AP), e.g., satellites, provide extensive coverage; meanwhile, it may introduce significant signal propagation delays due to the relatively long distances to the ground users, which can be multiple times longer than the packet durations in task-oriented communications. This phenomena is modeled as a new ``ripple effect'', which introduces spatiotemporally correlated interferences in SAGUIN. This paper studies the task scheduling problem in SAGUIN with ripple effect, and formulates it as a Markov decision process (MDP) to jointly minimize the age of information (AoI) at users and energy consumption at APs. The obtained MDP is challenging due to high dimensionality, partial observations, and dynamic resource constraints caused by ripple effect. To address the challenges of high dimensionality, we reformulate the original problem as a Markov game, where the complexities are managed through interactive decision-making among APs. Meanwhile, to tackle partial observations and the dynamic resource constraints, we adopt a modified multi-agent proximal policy optimization (MAPPO) algorithm, where the actor network filters out irrelevant input states based on AP coverage and its dimensionality can be reduced by more than an order of magnitude. Simulation results reveal that the proposed approach outperforms the benchmarks, significantly reducing users' AoI and APs' energy consumption.