Safe Adaptive Policy Transfer Reinforcement Learning for Distributed Multiagent Control.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Robotics

Background:

• Multiagent reinforcement learning (RL) training is challenging due to agent interference and safety constraints.
• Existing methods struggle with efficient knowledge transfer and adaptive learning in cooperative multiagent systems.

Purpose of the Study:

• To propose a safe adaptive policy transfer RL approach for multiagent cooperative control.
• To enhance learning efficiency and safety in multiagent systems through knowledge transfer.

Main Methods:

• Introduced a pioneer and follower off-policy policy transfer learning (PFOPT) method.
• Enabled transfer of policy representation and sample experience from a pioneer to follower agents.
• Utilized Wasserstein distance for adaptive adjustment of learning weights between prior experience and exploration.

Main Results:

• Trained distributed agents successfully completed collaborative tasks, maximizing rewards while minimizing constraint violations.
• Demonstrated satisfactory performance in learning speed and success rate compared to baseline methods.
• The PFOPT method effectively transferred knowledge and adapted learning based on policy distribution differences.

Conclusions:

• The proposed safe adaptive policy transfer RL approach significantly improves multiagent cooperative control.
• PFOPT offers an efficient and safe solution for complex multiagent training scenarios.
• This method provides a robust framework for knowledge sharing and adaptive learning in decentralized systems.