Dual-Policy Fusion for Multitask Multiagent Reinforcement Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Robotics

Background:

• Multiagent reinforcement learning (MARL) excels in cooperative tasks but struggles with dynamic, multi-task environments.
• Existing multitask MARL methods face challenges with negative transfer due to conflicting task knowledge.

Purpose of the Study:

• To introduce Dual-Policy Fusion for Multitask MARL (DPF-MTMARL) to improve adaptability and mitigate negative transfer.
• To develop an efficient training method for task-specific policies within DPF-MTMARL.
• To derive theoretical conditions for policy decentralization and enforce them via regularization.

Main Methods:

• DPF-MTMARL integrates a shared policy for common knowledge and task-specific policies for unique information.
• A novel learning method is proposed for efficient training of task-specific policies.
• Theoretical conditions for policy decentralization are derived and enforced using a regularization term.

Main Results:

• DPF-MTMARL significantly outperforms state-of-the-art baselines on both homogeneous and heterogeneous task sets.
• The proposed method effectively mitigates negative transfer in multitask MARL scenarios.
• Robust multitask learning capabilities were demonstrated.

Conclusions:

• DPF-MTMARL offers a robust solution for multitask MARL by effectively balancing shared and specific knowledge.
• The method enhances adaptability and performance in complex, dynamic environments.
• Theoretical analysis supports the practical implementation and decentralization of the joint policy.