Title:
Deep Learning for Dynamical Systems: Modeling, Prediction, and Control
Deep Learning for Dynamical Systems: Modeling, Prediction, and Control
Author(s)
Saha, Priyabrata
Advisor(s)
Mukhopadhyay, Saibal
Editor(s)
Collections
Supplementary to
Permanent Link
Abstract
Modeling and control of dynamical systems are fundamental problems across several scientific and engineering disciplines. Traditionally, dynamical systems or processes are modeled by a set of differential equations constructed based on physical principles and extensive experiments. However, in many practical scenarios, analytical modeling is very challenging and/or can only describe the system behavior partially. Furthermore, real-world dynamical processes are inherently nonlinear which makes the downstream task of control notoriously difficult. In recent years, the success of deep learning in various complex tasks has motivated many researchers to exploit deep learning for automatic modeling and control synthesis for dynamical systems from data. However, deep learning methods have their own drawbacks including high sample complexity, requirements of regular data structure, difficulty in long-term prediction, lack of generalizability outside of training scenarios, etc.
This thesis introduces a novel framework that leverages deep learning within the traditional identify-then-design paradigm to develop control policies for unknown or partially known dynamical systems, addressing the aforementioned challenges. The model learning process incorporates existing scientific knowledge and numerical techniques to facilitate learning from limited and partial observations as well as to improve long-term prediction accuracy and generalizability under system parameter changes. Given an identified or learned model, the control learning process leverages a self-supervised formulation guided by a Lyapunov-constrained deep neural network to ensure stability and improve sample efficiency. The proposed framework is evaluated in prediction and control tasks for several dynamical systems including multi-agent dynamics and spatiotemporal dynamics.
Sponsor
Date Issued
2023-07-30
Extent
Resource Type
Text
Resource Subtype
Dissertation