Learning to Accelerate Traffic Allocation Over Large-Scale Networks

Zhaoxing Yang, Guiyun Fan, Anjie Cao, Yuchen Guo, Wenlong Li, Tianyuan Liu, Haiming Jin

Abstract: In network systems with large-scale nodes and links, how to allocate traffic in real-time when both data transmission demands and link connections vary dynamically over time, in order to maximize the network's long-term total throughput under link capacity constraints, is a fundamental problem. However, state-of-the-art (SOTA) works lack scalability due to the requirement of solving the time-consuming constrained optimization problems online. Further, their reliance on reinforcement learning algorithms makes them inefficient to explore and train when facing systems with large-scale nodes and complex topologies. In this paper, we formulate such a challenging problem as the Dynamic Traffic Allocation (DTA) problem, which is generic and applicable across systems. Then, we propose the Fast Networked Control (FNC) policy framework for DTA, which removes the requirement of solving constrained optimization problems online, and only utilizes basic operations such as normalizations and comparisons to achieve fast constraints satisfaction. Further, we propose an imitation learning-based algorithm for efficiently optimizing the FNC policy. Experiments in large-scale networks show that our FNC policy achieves a maximal 10% improvement in demands satisfaction and 60x reduction in latency at the same time versus SOTA works.

External IDs:dblp:conf/infocom/YangFCGLLJ25