Event-Triggered ADP for Nonzero-Sum Games of Unknown Nonlinear Systems

Area of Science:

• Control Theory
• Artificial Intelligence
• Game Theory

Background:

• Nonlinear systems and nonzero-sum (NZS) games present challenges for control and optimization.
• Reinforcement learning (RL) and adaptive dynamic programming (ADP) offer iterative solutions for approximating optimal policies.
• Existing methods often rely on time-triggered updates, leading to potential inefficiencies.

Purpose of the Study:

• To develop an event-triggered ADP approach for NZS games in continuous-time nonlinear systems with unknown dynamics.
• To achieve approximate Nash equilibrium solutions efficiently.
• To reduce the computational and communication load compared to traditional methods.

Main Methods:

• Utilized critic neural networks to estimate value functions and actor neural networks for control policies.
• Implemented an event-triggered mechanism that updates system parameters only when specific conditions are met.
• Analyzed system stability and weight convergence under the proposed event-triggered framework.

Main Results:

• The event-triggered ADP algorithm successfully approximates Nash equilibrium solutions for NZS games.
• System stability and neural network weight convergence are guaranteed under mild assumptions.
• Significant reduction in communication and computation resource utilization was observed.
• The proposed method excludes Zeno behavior by ensuring a minimum inter-event time (MIET).

Conclusions:

• The event-triggered ADP approach is effective for NZS games of continuous-time nonlinear systems.
• This method offers a more resource-efficient alternative to time-triggered control.
• The theoretical guarantees of stability and convergence are maintained, along with practical feasibility.