Predictive Networked Control: From LLM‑Driven Context to Optimal Decentralization by Dr Tongxin Li

Abstract

This talk addresses distributed control of large networked systems under dynamic conditions and communication constraints. It introduces InstructMPC, which uses large language models to convert human‑provided context into predictive disturbance trajectories, and PredSLS, a framework that jointly integrates prediction errors and communication limits to achieve system‑level optimal control. A regret analysis reveals a non‑monotonic trade‑off between control performance and communication range, showing that increased communication can degrade performance when prediction errors propagate. The talk concludes by demonstrating how PredSLS identifies optimal localized communication neighborhoods for co‑designing controllers and network topology.

About the Speaker

Tongxin Li is an Assistant Professor and Presidential Young Fellow in the School of Data Science at The Chinese University of Hong Kong, Shenzhen. He received his PhD in CMS from the California Institute of Technology in 2022. His research focuses on control, online algorithms, and AI in power systems. He previously interned as an applied scientist at AWS Security and is a recipient of the SIGEnergy Doctoral Dissertation Award (Honorable Mention).