Safe Reinforcement Learning for Nonlinear Multiagent Systems Based on Min-Max DMPC

Area of Science:

• Control Theory
• Artificial Intelligence
• Robotics

Background:

• Nonlinear multiagent systems (MASs) require robust control strategies to handle disturbances and ensure safety.
• Traditional robust distributed model predictive control (DMPC) can be overly conservative due to fixed disturbance bounds and precise model requirements.

Purpose of the Study:

• To develop a safe reinforcement learning (RL) framework for nonlinear MASs.
• To enhance the performance and safety of control strategies by mitigating the conservatism of traditional robust DMPC.
• To ensure recursive feasibility and closed-loop stability during the adaptive learning process.

Main Methods:

• Integration of min-max DMPC as a baseline for robust control strategy generation.
• Application of safe RL to adaptively update controller parameters and disturbance sets online.
• Formal guarantee of recursive feasibility for the DMPC algorithm throughout the learning phase.
• Theoretical analysis of closed-loop stability.

Main Results:

• The proposed safe RL framework successfully mitigates the conservatism of robust DMPC.
• Adaptive online updates of controller parameters and disturbance sets are achieved.
• Recursive feasibility of the DMPC algorithm is formally guaranteed during learning.
• Theoretical stability analyses confirm the robustness of the approach.

Conclusions:

• The developed framework offers a safe and effective approach for controlling nonlinear MASs.
• The method enhances performance by adaptively managing uncertainties and ensuring stability.
• Validated through simulations, the approach demonstrates effectiveness and scalability for complex systems.