Title: Decentralized Learning of Interdependencies for Multi-Component Dynamic Systems

Date: 10th Nov 2026

Time: 10:00 to 12:00 EST

Location: Virtual via Zoom — https://gatech.zoom.us/j/98959804477?pwd=Evx56C26tohvh3a6Q4HQzyA2861uAz.1 (Meeting ID: 989 5980 4477,  Passcode: 070732)

Ayush Mohanty

Machine Learning PhD Student

H. Milton Stewart School of Industrial and Systems Engineering

Georgia Institute of Technology

Committee

1 Prof. Nagi Gebraeel (Advisor), H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology

2 Prof. Pascal Van Hentenryck, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology

3 Prof. Yao Xie, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology

4 Prof. Aaron Stebner,  George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology

5 Prof. Jeff Shamma, The Department of Industrial and Enterprise Systems Engineering, University of Illinois Urbana-Champaign

Abstract

Modern industrial systems comprise multiple interconnected subsystems whose temporal dynamics are often interdependent, making it essential to understand how changes in one subsystem influence others. Learning these interdependencies is particularly challenging in decentralized environments where raw data cannot be shared, and individual subsystems operate proprietary models that cannot be modified. This dissertation develops decentralized learning frameworks for characterizing cross-client temporal interdependencies in both linear and nonlinear dynamical systems while preserving the autonomy of individual client models. For linear systems, we establish a federated Granger causality framework that enables the discovery of unknown cross-client interdependencies. We further use this federated Granger causality framework to investigate the propagation of estimation uncertainty and its implications for interdependency learning. For nonlinear systems, the dissertation introduces graph attention-based learning with a predefined candidate graph structure, establishing theoretical relationships between learned attention coefficients and Jacobian-based interdependencies. It further develops a complementary framework for learning nonlinear interdependencies without requiring a predefined graph skeleton, where the downstream task of root cause analysis is explored. The proposed methods are evaluated on synthetic and industrial datasets, demonstrating their utility for decentralized prediction and interdependency characterization.