Asynchronous learning for actor-critic neural networks and synchronous triggering for multiplayer system.

Area of Science:

• Control Theory
• Artificial Intelligence
• Game Theory

Background:

• Multiplayer nonzero-sum differential games present significant challenges in finding optimal strategies.
• Traditional methods often struggle with adaptability and computational efficiency in complex, dynamic environments.

Purpose of the Study:

• To develop a novel asynchronous learning algorithm for adaptive solution of Nash equilibrium in multiplayer nonzero-sum differential games.
• To reduce communication overhead while maintaining learning accuracy and stability.

Main Methods:

• Utilizing an actor-critic neural network structure and reinforcement learning.
• Implementing distributed asynchronous policy iteration with event-induced updates.
• Employing a central event generator to manage communication burden.

Main Results:

• The algorithm ensures closed-loop asymptotic stability and uniform ultimate convergence.
• Demonstrated significant reduction in sampling numbers on a four-player nonlinear system without compromising learning accuracy.
• Successfully applied to aircraft lateral-directional stability and nonlinear vehicle adaptive cruise control.

Conclusions:

• The proposed asynchronous learning algorithm effectively solves Nash equilibrium in multiplayer nonzero-sum differential games.
• The framework offers an efficient and accurate adaptive control strategy for complex systems.
• The approach is versatile, applicable to various control and stability problems.