GNN Model for Time-Varying Matrix Inversion With Robust Finite-Time Convergence.

Area of Science:

• Computational mathematics
• Artificial intelligence
• Neural networks

Background:

• Gradient neural networks (GNNs) excel at static matrix inversion but struggle with time-varying matrices, often yielding residual errors, especially under noisy conditions.
• Existing zeroing neural networks (ZNNs) for time-varying inversion are complex, require matrix derivative information, and involve inherent inversion operations on digital computers.

Purpose of the Study:

• To propose a unified gradient neural network (GNN) model capable of handling both static and time-varying matrix inversion.
• To achieve finite-time convergence and enhanced robustness to noise in time-varying matrix inversion tasks.

Main Methods:

• Development of a novel, unified GNN model designed for both static and dynamic matrix inversion.
• Theoretical analysis to demonstrate finite-time convergence properties under specified conditions, including the presence of bounded noise.
• Comparative simulations against established GNN and ZNN models for time-varying matrix inversion.

Main Results:

• The proposed unified GNN model demonstrates finite-time convergence for time-varying matrix inversion, even with bounded noise.
• Simulations confirm superior convergence speed compared to existing GNN and ZNN models.
• The new GNN model exhibits enhanced robustness against noise in time-varying matrix inversion tasks.

Conclusions:

• The unified GNN model provides a simpler and more effective approach for both static and time-varying matrix inversion.
• This model overcomes limitations of traditional GNNs and ZNNs, offering improved performance and noise resilience.