Theory Guided Data-Driven Method for Scalable Scheduling in Time-Sensitive Network

Time-Sensitive Networking (TSN) has emerged as a vital networking paradigm for timely and reliable data transmission. Recently, researchers are increasingly focusing on the scalability of algorithms for scheduling problems over diverse network topologies and flows based on deep reinforcement learning methods (DRL). However, the lack of reasonable characterization of flow conflicts and dependencies leads to limited generalization and schedulability under the discrepancies in characteristics and number of flows. To address those issues, we propose a data-driven scheduling method guided by flow sequence conflict theory, which improves generalization and schedulability over different flow characteristics and topologies. Specifically, we design a network feature encoding scheme to formulate a reasonable characterization of flow conflicts by integrating a graph-based flow sequence representation model with graph attention networks (GATs). Regarding the abovementioned characterization as the acknowledge embedding, a data-driven method is designed to learn two sub-policies in the scheduling process, including flow selection and solution searching for each selected flow. Within the scheduling process, a multi-head attention-based neural network is designed to generate the scheduling sequence of variable length, which characterizes the dependencies between flows with different numbers. Simulation results demonstrate that compared to existing DRL-based methods, the runtime of scheduling at scale is increased by at least 100 % in complex traffic scenarios while the slot utilization on the links is also improved compared to other methods.