Self-triggered adaptive dynamic programming based on experience-replay and spectral adaptive law

Area of Science:

• Control Systems Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Optimal control of unknown nonlinear systems presents significant challenges.
• Existing adaptive dynamic programming (ADP) methods often require continuous monitoring and struggle with generalization.
• Event-triggered approaches necessitate constant state observation for control triggering.

Purpose of the Study:

• To propose a novel self-triggered adaptive dynamic programming (ADP) framework.
• To enhance the robustness and efficiency of ADP for unknown nonlinear systems.
• To reduce communication overhead in control systems through a self-triggered mechanism.

Main Methods:

• Integration of experience replay (ER) and a spectral adaptive law (SPAL) within the ADP framework.
• Development of a new critic neural network (NN) weight updating law based on SPAL for improved generalization.
• Implementation of a self-triggered mechanism to predict the next triggering moment, avoiding continuous state monitoring.

Main Results:

• The proposed SPAL-based NN weight updating law enhances generalization ability for optimal control.
• The ER-based ADP method, utilizing a SPAL-based NN system identifier, demonstrates improved robustness.
• The self-triggered mechanism effectively conserves communication compared to event-triggered methods.

Conclusions:

• The novel self-triggered ADP framework effectively achieves optimal control for unknown nonlinear systems.
• The integration of ER and SPAL leads to enhanced robustness and efficiency.
• The proposed method significantly reduces communication requirements, validated through simulation.