Core Concepts
Smoothed online learning algorithms enable efficient prediction in piecewise affine systems.
Abstract
The content discusses the application of smoothed online learning algorithms for prediction in piecewise affine systems. It introduces a new framework for prediction and simulation in such systems, focusing on regret minimization and efficient optimization. The paper presents algorithms and guarantees for one-step prediction and multi-step simulation regret, emphasizing the importance of directional smoothness in achieving low regret rates. Technical tools and proofs are provided to support the proposed methods.
Structure:
Introduction: Discusses the significance of planning through piecewise-affine systems.
Setting: Describes the problem setting for online PWA regression.
Algorithm and Guarantees:
Regret for One-Step Prediction in PWA Systems.
Guarantees for Simulation Regret.
Analysis:
Parameter Recovery: Details the control of regret due to parameter estimation errors.
Mode Prediction: Addresses challenges in mode classification stability using surrogate loss functions.
Stats
"Our solution takes inspiration from, and establishes a connection between, two rather different fields." (Abstract)
"Regret does not require uniform identification of system parameters, which is typically impossible." (Abstract)
"Discontinuities and non-smoothness pose significant challenges to planning and modeling contact-rich dynamics." (Discussion)