Event-Triggered Multigradient Recursive Reinforcement Learning Tracking Control for Multiagent Systems.

Area of Science:

• Control Systems Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Nonlinear multiagent systems (MASs) present complex control challenges.
• Distributed reinforcement learning offers a promising approach for MAS control.
• Existing methods may suffer from local optima and high update frequencies.

Purpose of the Study:

• To investigate event-triggered multigradient recursive reinforcement learning for nonlinear MASs.
• To enhance energy conservation in MASs through optimized control strategies.
• To ensure the stability and boundedness of signals within the MAS.

Main Methods:

• Utilizing a critic neural network (NN) for utility function estimation.
• Employing an actor NN to approximate uncertain system dynamics.
• Implementing a multigradient recursive (MGR) strategy for NN weight vector learning.
• Applying an event-triggered mechanism to reduce controller updates.

Main Results:

• The proposed MGR strategy overcomes local optima and reduces initial value dependence.
• Reinforcement learning and event-triggered mechanisms improve energy conservation.
• Lyapunov theory proves that all signals in the MAS are semiglobal uniformly ultimately bounded (SGUUB).
• Simulation results validate the effectiveness of the developed control strategy.

Conclusions:

• The event-triggered multigradient recursive reinforcement learning approach is effective for nonlinear MASs.
• The strategy ensures system stability and improves energy efficiency.
• This method offers a robust alternative to traditional gradient descent techniques in MAS control.