Reliability-oriented dynamic task offloading for time-varying satellite IoT: A lightweight multi-agent cooperative strategy optimization method

Abstract

The integration of satellite Internet of Things (IoT) with Mobile Edge Computing (MEC) has revolutionized global connectivity, enabling intelligent applications in remote regions while presenting distinctive reliability challenges stemming from intermittent satellite connections and spatio-temporal workload dynamic. However, traditional centralized approaches fail to address these challenges, as their reliance on stable connections and centralized control fundamentally conflicts with the dynamic nature of satellite networks. Motivated by these challenges, we propose a decentralized cooperative task offloading framework where satellite edge servers independently manage task oddloading, inter-satellite task migration, and resource allocation. To enhance offloading reliability in bursty traffic scenarios, we leverage Stochastic Network Calculus (SNC) to integrate communication and computation failure probabilities into our optimization framework. Aiming at coordinated decision-making and global optimization, this work presents a lightweight cooperative task offloading algorithm (MA-LWCTO) utilizing multi-agent soft actor–critic, where the satellites make decisions through local state and shared neighboring policy information. In response to the challenges of increasing computational complexity and state space expansion, we develop an information extraction mechanism based on long short-term memory variational auto-encoder with attention (VLAEA), facilitating efficient information while reducing communication overhead. Extensive simulations based on pre-obtained satellite trajectory data demonstrate that the proposed algorithm significantly enhances reliability while reducing system service costs in satellite–terrestrial MEC environments.