Lateral Transfer Learning for Multiagent Reinforcement Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Robotics

Background:

• Existing multiagent reinforcement learning (MARL) transfer learning methods are limited to homogeneous agents or similar tasks.
• Cross-task transfer in multiagent systems faces challenges like negative transfer and domain specificity.

Purpose of the Study:

• To propose a versatile cross-task transfer learning method, multiagent lateral transfer (MALT), for MARL.
• To alleviate the training burden in MARL by enabling knowledge reuse across dissimilar tasks and agent types.

Main Methods:

• MALT utilizes features as the transfer object, inspired by progressive networks.
• Pretrained policy networks are assigned using clustering, and an attention module enhances the transfer framework.
• The method imposes no strict requirements on source or target tasks, accommodating heterogeneous agents and diverse scenarios.

Main Results:

• MALT successfully transfers knowledge among heterogeneous agents, avoiding negative transfer even in vastly different tasks.
• Experimental results across various scenarios demonstrate MALT's flexibility in cooperative, competitive, homogeneous, and heterogeneous MARL settings.
• The proposed method significantly outperforms baseline approaches in MARL knowledge transfer.

Conclusions:

• MALT represents a pioneering, all-purpose cross-task transfer learning solution for MARL.
• The method offers a flexible and effective approach to knowledge reuse, applicable to a wide range of MARL configurations.
• MALT addresses key limitations of previous transfer learning techniques in multiagent systems.