Communication networks are the infrastructure of information era and systems of stunning complexity. Wireless networks that span entire cities and wired networks that span continents perform a sort of alchemy. They transform unreliable electromagnetic signals into reliable communication streams. We are currently pushing against the limits of our ability to design and operate communication systems. The TILOS team will develop the foundations of intelligent autonomous networks that operate with minimal human supervision.
At their core, autonomous networks are optimization systems made up of transceivers that sense the environment and decide on resource allocations that are beneficial to the network as a whole. Optimization in communication networks presents several challenges. Problems involve large numbers of discrete and continuous variables over nonconvex sets and metrics induced by interference. Solutions must be obtained by distributed methods, operate in real time, and be robust to inaccurate and partial sensing.
To address these challenges, we will work with foundation team to deploy sequential sampling and deep learning methods. The TILOS team will address these challenges by focusing on three technical innovations:
Multiscale optimization to enable scalability and real time operation. We propose to use AI-enabled decentralized and stochastic optimization of a large number of discrete and continuous (random) variables, dynamically, with nonconvex constraints, over multiple time scales. A unique challenge is that distributed real-time optimization of networks must meet the end goal of increased information transfer rate. Extending our prior work on active learning and data acquisition, a key challenge that we will tackle is how to balance the learning/optimization gains against communication and message passing overhead. Our work will naturally complement addressing communication vs. convergence in federated learning, and provide information-theoretic fundamental limits on overheads associated with near real-time learning and optimization.
Autonomous tuning to enable distributed implementation and robustness. Large-scale distributed optimization can be used to design multi-scale network protocols. However, the solutions are often followed by a fine-tuning step in which human network managers use heuristics to optimize architecture-level hyperparameters. We propose to capitalize on the abundance of the data and past design instances via empirical (black box) optimization of architecture-level hyperparameters.
Integration of Expert Knowledge
Integration of expert knowledge and physics to leverage existing engineered solutions. Automated network design and optimization often rely on domain knowledge to build statistical and geometric models, and to guide empirical optimization steps. Indeed, early generations of wireless communication systems relied on human experts and their intuition about (simplified) physics-based models of wireless channels. This leads to two key research directions. First, we will apply techniques form machine teaching to directly capture human intuition along with the physics of propagation. We will also pursue a next generation of AI-enabled optimization techniques that are able to integrate physics-based models studied in signal processing, communications, and information theory.