Computation offloading and resource allocation in satellite edge computing networks: A multi-agent reinforcement learning approach

The development of Low Earth Orbit (LEO) satellite networks and Mobile Edge Computing (MEC) technologies supports the placement of MEC servers on LEO satellites, facilitating computation offloading in remote areas where computing resources are limited. However, the on-board computing and communication resources of LEO satellite networks are similarly constrained, while the system environment remains highly dynamic and complex. Moreover, diverse task requirements often require offloading across multiple time slots, which increases the complexity of offloading decisions and resource allocation for terrestrial tasks. In this study, we model this problem within satellite edge computing networks as a partially observable Markov decision process (POMDP). To achieve effective joint optimization, we introduce a multi-agent recurrent attentional double delayed deep deterministic policy gradient (MARATD3) algorithm. First, we utilize the recurrent neural network (RNN) to summarize historical observations of users, which improves adaptability to dynamic system environments and enables accurate predictions of system states. Then, a multi-head attention mechanism is introduced to strengthen the ability of user agents to capture critical information within the joint state space, reduce interference from irrelevant information, and improve training efficiency. According to the experimental results, MARATD3 achieves a considerable reduction in energy consumption and delay relative to the baseline algorithms while maintaining task delay and resource constraints.