Resilient Supervisory Multi-Agent Systems

Area of Science:

• Computer Science
• Artificial Intelligence
• Network Science

Background:

• Multi-agent systems (MAS) are vulnerable to disruptions in central coordination, termed 'leader decapitation'.
• Restoring normal operation after such failures is critical for system reliability.

Purpose of the Study:

• To develop a methodology for enhancing the resilience of multi-agent networks against coordination function failure.
• To enable timely restoration of operational normalcy using machine learning.

Main Methods:

• Agents are equipped with independent learning modules enabling role discovery within the system's coordinating strategy.
• Machine learning algorithms facilitate autonomous strategy implementation when central coordination ceases.
• Agents incrementally identify system task specifications and optimize individual strategies for the common goal.

Main Results:

• Demonstrated a methodology for creating resilient multi-agent systems.
• Showcased the capability of agents to autonomously adapt and maintain system function post-decapitation.
• Validated the effectiveness of machine learning in decentralized coordination restoration.

Conclusions:

• The proposed machine learning approach significantly enhances multi-agent network resilience to leader decapitation.
• Autonomous role discovery and strategy optimization by agents ensure operational continuity.
• This methodology offers a robust solution for maintaining system functionality in decentralized networks.