Online Learning for Provable Adaptive Nonlinear Control

Abstract

Control theory research, as well as its applications, have seen a resurgence in recent years with the riseof machine learning and deep learning in particular. However, machine learning based approaches fornonlinear control lack rigorous theoretical guarantees that are important for systems with strict safetyand performance requirements.The goal of this proposal concerns a novel machine learning viewpoint to control in an effortto add new methods with provable finite-time performance guarantees for more robust settings thanclassical theory. These new methods are based on recent developments in online learning coupledwith novel convex relaxations, giving rise to the methodology of nonstochastic control theory. Thelatter has been successfully applied to linear dynamical systems, giving new algorithms with provableguarantees.The crux of this research proposal is to provide novel methods with provable guarantees fornonlinear control. We propose several research directions based on the decision making frameworkof online convex optimization, including the use of policy regret as a performance metric, and thetechniques developed therein. The directions span adaptive policy regret guarantees for time varyingsystems based on Lyapunov#s linearization method, a novel boosting approach to control, and deepneural network control methods with provable guarantees.

Document Details

Document Type
DoD Grant Award
Publication Date
Jan 12, 2023
Source ID
N000142312156

Entities

People

  • Elad Hazan

Organizations

  • Office of Naval Research
  • Trustees of Princeton University
  • United States Navy

Tags

Fields of Study

  • Computer science

Readers

  • Adaptive Control and Estimation with Uncertainty in Dynamic Systems.
  • Agent-Based Social Robotics and Mobile-Assisted Learning in Virtual Environments.

Technology Areas

  • AI & ML
  • AI & ML - Machine Learning Algorithms
  • AI & ML - Neural Networks