AUTO: Adaptive Congestion Control Based on Multi-Objective Reinforcement Learning for the Satellite-Ground Integrated Network

The satellite-ground integrated network is highly heterogeneous with diversified applications. It requires congestion control (CC) to achieve consistent high performances in both long-latency satellite networks and large-bandwidth terrestrial networks and cope with different application requirements. However, existing schemes can hardly achieve these goals, for they cannot balance the objectives of CC (i.e., throughput, delay) adaptively and are not objective-configurable. To address these limitations, we propose and implement a novel adaptive CC scheme named AUTO, based on Multi-Objective Reinforcement Learning (MORL). It is environment-adaptive by training a MORL agent and a preference adaptation model. The first can generate optimal policies for all possible preferences (i.e., the relative importance of objectives). The latter automatically selects an appropriate preference for each environment, by taking a state sequence as input to recognize the environment. Meanwhile, AUTO can satisfy diversified application requirements by letting applications determine the input preference at will. Evaluations on emulated networks and the real Internet show that AUTO consistently outperforms the state-of-the-art in representative network environments and is more robust to stochastic packet loss and rapid network changes. Moreover, AUTO can achieve fairness against different CC schemes.