A Dynamic Neural Network-Based Control Method Using Reinforcement Learning for Nonlinear Parameter-Varying System...

Area of Science:

• Control Systems Engineering
• Artificial Intelligence
• Aerospace Engineering

Background:

• Nonlinear parameter-varying (NPV) systems present significant control challenges due to their dynamic nature.
• Existing control methods often struggle with adaptability and generalization across different system parameters.

Purpose of the Study:

• To propose a dynamic neural network (DNN)-based control method for optimal control of NPV systems.
• To enhance control policy generalization and adaptability to system parameter variations.
• To achieve efficient control without requiring extensive retraining or sample collection.

Main Methods:

• A DNN-based control policy (DNN-CP) with static shared layers and a parameter-related dynamic layer was constructed.
• An extreme learning machine (ELM) model predicted dynamic weights based on system parameters.
• A combined supervised pretraining and reinforcement learning (RL) algorithm was developed for training.
• Shared layers were optimized via constrained multiobjective problems, and the ELM model was tuned for parameter-specific objectives.

Main Results:

• The DNN-CP demonstrated effective generalization capabilities across different systems within the parameter space.
• The control policy could be applied immediately without sample collection or fine-tuning for new systems.
• The proposed method achieved superior control performance compared to existing approaches, especially for systems with continuously varying parameters.
• Validation was successfully performed on morphing aircraft applications.

Conclusions:

• The developed DNN-CP and training algorithm offer a robust solution for optimal control of NPV systems.
• The method significantly improves data efficiency and adaptability, enabling real-time application.
• This approach represents a breakthrough in achieving adaptive and generalized control for complex dynamic systems.