UNUM Framework for Enhanced Network Control

This technical report presents UNUM, a novel framework designed to improve network control tasks such as congestion control and adaptive bitrate streaming. The framework addresses the limitations of current state estimation methods, which often rely on instantaneous or running-average metrics, leading to imprecise approximations of the true network state. UNUM employs a unified network state embedder that utilizes the self-attention mechanism of Transformers and diverse training datasets to learn rich, latent state representations. The report details the architecture of UNUM, including its components like the Unum Collector, Embedder, and Predictor, and describes how these elements work together to enhance control performance. Experimental results demonstrate that integrating UNUM embeddings into existing network control frameworks significantly improves their effectiveness, as evidenced by performance metrics from real and synthetic settings. The report also discusses the implementation of key building blocks necessary for running UNUM in user-space, emphasizing the framework's potential for future enhancements and adaptability to various network conditions.