Multi-Agent Deep Reinforceinent Learning-Based Offloading Computation and Routing in Coope...

The increasing demand for tasks and dynamically changing loads in the Low Earth Orbit (LEO) satellite networks creates significant challenges in terms of computing and routing. Currently, LEO satellites primarily offload tasks to ground stations or satellites within their line of sight, failing to fully utilize the computational resources of the entire network. In addition, existing routing algorithms fail to consider on-satellite loads and computational capacities, leading to bottlenecks in network routing as some satellites with limited processing capacity become overwhelmed. In this paper, the tasks generated by the source satellite can be offloaded to either satellites or ground stations while routing to the destination satellite. The offloading computation and routing decision problems are investigated to minimize the maximum delay. To solve this challenging problem, we first convert the optimization variables, encompassing both routing and computation offloading, into a form that depends solely on the latter, and model the problem as the Markov Decision Process (MDP). Subsequently, the problem is addressed using an algorithm based on Multi-Agent Proximal Policy Optimization (MAPPO), where multiple agents cooperatively determine routing and offloading computation strategies. Simulation results show that the proposed scheme achieves better delay performance.

Multi-Agent Deep Reinforceinent Learning-Based Offloading Computation and Routing in Cooperative LEO Satellite Communication Network

Yunyi Yan, Ming Zeng, Zijian Yang, Zesong Fei, Beijing Institute of Technology