Experience-based integral reinforcement learning consensus for unknown multi-agent systems

Area of Science:

• Control Theory
• Artificial Intelligence
• Robotics

Background:

• Multi-agent systems (MAS) present complex control challenges, especially with unknown dynamics.
• Achieving consensus (agreement) in MAS is crucial for coordinated tasks.
• Traditional policy iteration methods often require system model identification.

Purpose of the Study:

• To develop an optimal consensus control strategy for nonlinear MAS with unknown dynamics.
• To implement a policy iteration algorithm using online integral reinforcement learning.
• To address and overcome the challenge of local optima in online learning.

Main Methods:

• A critic-actor neural network architecture was integrated into policy iteration.
• Online integral reinforcement learning was employed to handle unknown system dynamics.
• An experience-based weight-tuning law was introduced to ensure persistent excitation.

Main Results:

• The proposed algorithm successfully achieved optimal consensus control.
• The system demonstrated asymptotic stability.
• Neural network weights were shown to converge during the learning process.
• Simulation studies validated the algorithm's effectiveness and correctness.

Conclusions:

• The critic-actor neural network approach effectively bypasses the need for dynamics identification in MAS control.
• The developed algorithm provides a robust solution for optimal consensus in nonlinear systems.
• The findings contribute to advancements in intelligent control for multi-agent systems.