Robust backpropagation training algorithm for multilayered neural tracking controller

Area of Science:

• Control Systems Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Neural networks (NNs) are increasingly used in complex control systems.
• Ensuring NN convergence and minimizing tracking errors in dynamic environments remains a challenge.
• Disturbances can significantly degrade the performance of NN-based controllers.

Purpose of the Study:

• To develop a robust training algorithm for NN tracking control systems.
• To improve the convergence properties of multilayered NNs under disturbance.
• To minimize estimation and tracking errors in NN controllers.

Main Methods:

• A robust backpropagation training algorithm incorporating a dead zone scheme was developed.
• The algorithm was applied to a three-layered neural network with adjustable weights.
• Convergence analysis and proof were provided for the proposed algorithm.

Main Results:

• The dead zone scheme ensures NN convergence in the presence of disturbances.
• A smaller dead zone range directly correlates with reduced NN estimation error.
• Reduced estimation error leads to a smaller tracking error for the NN controller.

Conclusions:

• The proposed backpropagation algorithm with a dead zone is effective for online tuning of NN tracking control.
• The findings demonstrate that optimizing the dead zone parameter enhances control system accuracy.
• The algorithm's convergence proof and applicability can be extended to NNs with more hidden layers.