Meta-Learning Framework for Uncertain Nonlinear System Control

Researchers propose a meta-learning-based control framework for reference tracking in uncertain nonlinear systems, leveraging limited target system data by using offline data from structurally similar source systems. The framework adapts the implicit model-agnostic meta-learning (iMAML) algorithm to control settings, operating in two phases: an offline meta-training phase that learns an aggregated representation of shared dynamics from source data, and an online meta-adaptation phase that fine-tunes this representation on the target system. This approach aims to design optimal controllers efficiently when collecting data from the target system is challenging. The paper is published on arXiv with ID 2605.22513.