Finite time convergent learning law for continuous neural networks

Area of Science:

• Control Systems Engineering
• Machine Learning
• Nonlinear Dynamics

Background:

• Neural networks are employed for non-parametric modeling of uncertain systems.
• System uncertainties arise from external perturbations and incomplete knowledge of dynamics.
• Continuous learning methods often struggle with robustness and convergence in such scenarios.

Purpose of the Study:

• To design a discontinuous finite-time convergent learning law for neural networks.
• To develop an adaptive algorithm for adjusting neural network weights in uncertain environments.
• To improve the modeling accuracy and robustness of neural networks for systems with ODEs.

Main Methods:

• A discontinuous finite-time convergent learning law was designed for neural networks.
• A non-standard Lyapunov function was utilized for algorithm derivation.
• A nonlinear robust compensator was integrated to reject specific perturbations.

Main Results:

• The proposed learning law demonstrated improved performance compared to classical continuous methods.
• Numerical examples validated the effectiveness of the discontinuous learning algorithm.
• The methane production model showcased practical engineering benefits.

Conclusions:

• The discontinuous learning law offers enhanced modeling capabilities for uncertain systems.
• The adaptive algorithm provides a robust approach for neural network weight adjustment.
• This method shows significant potential for engineering applications requiring accurate system identification.