Learning-based Task Offloading in Dynamic Orbital Edge Computing Network

In existing satellite communication model (SCM), satellites receive remote commands from ground station, then downlink raw sensed data to ground station directly. To tackle the challenges of high delay, intermittent downlink and congested ground station, edge computing for SCM is proposed, in which the flying satellites can locally process raw data. In SCM, the computing resources are time-variant and difficult to be obtained accurately due to rapid mobility of satellites. Multi-armed bandit (MAB) is a well-known scheme that has great potential to achieve nearly optimal task offloading for edge computing in mobility cases. However, existing MAB based works are difficult to apply in SCM. First, the distance difference between each satellite may be very large and the existing works have ignored the distance factor in exploration. This may cause huge errors for exploration cost evaluation and inevitable offloading failures. Second, the exploration and exploitation may have several order of magnitude difference while the existing work only use a constant to adjust the trade-off between the exploration and exploitation. As a result, it may take a long learning process to balance the trade-off, causing low offloading efficiency. To address the above challenges, in this work, we propose DMUCB as a Distance-aware and Mobility-aware task offloading algorithm based on MAB theory. The offloading decisions are made without any prior knowledge of computing resources. First, DMUCB provide a novel exploration method by jointly considering the distance, task size and the accumulative selected times to achieve an accurate exploration cost. Second, the difference of exploration and the new exploitation is largely reduced by our well-designed normalization mechanism for fast converge to the optimal offloading. Third, we present a time-variant parameter to effectively adapt the dynamic satellite communication environment. Simulation results show that DMUCB achieves close-to-optimal task offloading delay performance for satellite communication.