Deep Multiagent Reinforcement Learning for Task Offloading and Resource Allocation in Satellite Edge Computing

As a supplement to terrestrial communication networks, satellite edge computing can break through geographical limitations and provide on-orbit computing services for people in some remote areas to achieve truly seamless global coverage. Considering time-varying channels, queue delays, and dynamic loads of edge computing satellites, we propose a multiagent task offloading and resource allocation (MATORA) algorithm with weighted latency as the optimization goal. It is a mixed integer nonlinear problem decoupled into task offloading and resource allocation subproblems. For the offloading subproblem, we propose a distributed multiagent deep reinforcement learning algorithm, and each agent generates its own offloading decision without knowing the prior knowledge of others. We show that the resource allocation problem is convex and can be solved using convex optimization methods. The experiment shows that the proposed algorithm can better adapt to the change of channel and the dynamic load of edge computing satellite, and it can effectively reduce task latency and task drop rate.