The paper presents a Lyapunov-based approach to derive weight adaptation laws for a deep residual neural network (ResNet)-based adaptive controller. The key highlights are:
Motivation: Deep neural network (DNN)-based controllers can compensate for unstructured uncertainties, but existing methods either use static DNN models or require offline training of inner-layer weights. This motivates the need for a ResNet-based adaptive controller with real-time weight adaptation.
Approach: The ResNet is expressed as a composition of building blocks involving a shortcut connection across a fully-connected DNN. A constructive Lyapunov-based approach is provided to derive weight adaptation laws for the ResNet using the gradient of each DNN building block.
Novelty: This is the first result on Lyapunov-derived adaptation laws for ResNets, which pose additional mathematical challenges compared to fully-connected DNNs due to the shortcut connections.
Analysis: A nonsmooth Lyapunov-based analysis is provided to guarantee asymptotic tracking error convergence. The analysis ensures the system state remains within a compact domain where the universal function approximation property of the ResNet holds.
Simulations: Comparative Monte Carlo simulations demonstrate that the ResNet-based adaptive controller provides a 64% improvement in tracking and function approximation performance compared to an equivalent fully-connected DNN-based adaptive controller.
Key Advantages: The ResNet architecture overcomes the vanishing gradient problem present in fully-connected DNNs, enabling faster weight adaptation and better compensation of system uncertainties.
To Another Language
from source content
arxiv.org
Principais Insights Extraídos De
by Omkar Sudhir... às arxiv.org 04-12-2024
https://arxiv.org/pdf/2404.07385.pdfPerguntas Mais Profundas