上海科技大学知识管理系统

In satellite-terrestrial integrated networks, it is a common practice to distribute real-time tasks from low Earth orbit (LEO) satellites to ground stations (GSs) for data processing. However, it remains an open problem how to match tasks with proper GSs in an online fashion with unknown dynamics, e.g., transmission latency. Moreover, such a problem is further complicated by the non-trivial interaction between the decision-making procedure and long-term constraints on time-averaged energy consumptions. In this paper, by formulating the energy-constrained online matching problem with unknown transmission latency as a constrained Combinatorial Multi-Armed Bandit (CMAB) problem, we adopt bandit learning methods and virtual queue techniques to deal with the exploration-exploitation tradeoff and long-term constraints, respectively. With an effective integration of online learning and online control, we propose a Task-matching and Resource-allocation with Data-driven Bandit Learning (TRDBL) scheme. Our theoretical analysis shows that TRDBL achieves a sublinear regret bound with a time-averaged energy constraints guarantee in the long run. Through simulation results we not only verify our theoretical analysis but also demonstrate the outperformance of TRDBL in terms of both task latency reduction and energy efficiency. © 2021 IEEE.