Standard representation and unified stability analysis for dynamic artificial neural network models.

Area of Science:

• Control Engineering
• Artificial Intelligence
• Nonlinear System Dynamics

Background:

• Dynamic artificial neural network (DANN) models are crucial for addressing complex nonlinear dynamical system identification and control challenges.
• Existing stability conditions for DANNs can be overly conservative, limiting their practical application.
• Understanding the approximation capabilities and error bounds of DANNs is essential for reliable system modeling.

Purpose of the Study:

• To provide an overview of popular DANN models and their mathematical representations.
• To develop novel, less conservative convex stability conditions for DANNs.
• To characterize nonlinear dynamical systems that can be universally approximated by DANNs.

Main Methods:

• Description of three popular DANN classes, including their architectures and block diagram transformations.
• Characterization of universally approximable nonlinear dynamical systems with rigorous error bounds.
• Development of a unified framework using linear matrix inequality (LMI)-based conditions for stability analysis.

Main Results:

• Proposed convex stability conditions that are less conservative than previous results.
• Inclusion of additional information like local slope restrictions and odd nonlinearities in stability analysis.
• Demonstration of reduced conservatism through a theoretical example.

Conclusions:

• The proposed stability conditions offer improved performance and reduced conservatism for DANNs.
• The unified framework facilitates stability and performance analyses across different DANN classes.
• This work advances the application of DANNs in nonlinear system identification and control.