Two-Layer Reinforcement Learning for Output Consensus of Multiagent Systems Under Switching Topology

Area of Science:

• Control Systems Engineering
• Artificial Intelligence
• Networked Systems

Background:

• Achieving consensus in multiagent systems is crucial for coordinated behavior.
• Switching topologies in multiagent systems present challenges for traditional control algorithms due to time-varying system dynamics.
• Existing reinforcement learning methods struggle with time-varying kernel matrices inherent in switching topologies.

Purpose of the Study:

• To develop a data-based reinforcement learning algorithm for output consensus in discrete-time multiagent systems with switching topology.
• To overcome the limitations of existing algorithms when dealing with switching-varying kernel matrices.
• To propose a distributed control policy applicable to both fixed and switching topologies.

Main Methods:

• A novel two-layer reinforcement learning algorithm is proposed to handle the switching-varying kernel matrix.
• A data-based distributed control policy is designed for implementation.
• Convergence analysis of the proposed algorithm is performed.

Main Results:

• The proposed algorithm effectively achieves data-based output consensus under switching topology.
• The developed control policy is applicable to both fixed and switching network topologies.
• The method removes restrictive assumptions on the leader's dynamic matrix eigenvalues found in prior work.

Conclusions:

• The novel two-layer reinforcement learning approach successfully addresses the challenge of output consensus in discrete-time multiagent systems with switching topology.
• The data-based distributed control policy offers a practical solution for real-world applications.
• Simulation examples validate the algorithm's effectiveness and convergence properties.