Reinforcement learning-based funnel control and privacy preservation for multi-agent systems with input dead-zone.

Area of Science:

• Control Systems Engineering
• Artificial Intelligence
• Cybersecurity

Background:

• Multi-agent systems (MAS) often face challenges with input dead-zone constraints and communication burdens.
• Ensuring data privacy during state information transmission is critical in distributed control systems.
• Adaptive control strategies are needed to handle uncertainties and nonlinearities in complex systems.

Purpose of the Study:

• To design a privacy-preserving reinforcement learning-based funnel controller for MAS with input dead-zone constraints.
• To guarantee tracking errors remain within prescribed boundaries despite system uncertainties.
• To develop an efficient, secure data-exchange mechanism for MAS control.

Main Methods:

• An adaptive funnel controller was formulated using an actor-critic reinforcement learning paradigm.
• Fuzzy logic was employed to approximate uncharacterized system nonlinearities.
• An event-triggered scheme was introduced for efficient control signal updates.
• The Paillier cryptographic scheme was integrated for secure data exchange.

Main Results:

• The proposed controller successfully guaranteed that tracking errors stayed within predefined limits.
• The event-triggered scheme effectively reduced communication load while maintaining control performance.
• The cryptographic mechanism ensured the privacy of state information during transmission.
• Simulations validated the controller's feasibility and effectiveness in handling input dead-zone constraints.

Conclusions:

• The developed strategy offers a robust and secure solution for controlling multi-agent systems with input dead-zone constraints.
• The integration of reinforcement learning, fuzzy logic, and cryptography enhances system performance and data security.
• This approach provides a foundation for advanced, privacy-aware control in complex distributed systems.