Interaction Pattern Disentangling for Multi-Agent Reinforcement Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Reinforcement Learning

Background:

• Deep cooperative multi-agent reinforcement learning excels in complex control tasks.
• Current methods often overfit due to intertwined entity interactions and noisy data.

Purpose of the Study:

• To introduce a novel method, interactiOn Pattern disenTangling (OPT), for disentangling entity interactions.
• To improve generalizability and interpretability in multi-agent reinforcement learning by filtering noisy interactions.

Main Methods:

• OPT disentangles entity interactions into prototypes representing underlying patterns within subgroups.
• A sparse disagreement mechanism encourages prototype sparsity and diversity.
• An aggregator with learnable weights restructures prototypes into a compact interaction pattern.
• Mutual information maximization addresses training instability under partial observability.

Main Results:

• Experiments demonstrate OPT's superiority over state-of-the-art methods on single-task, multi-task, and zero-shot benchmarks.
• The method effectively filters noisy interactions, enhancing model performance.

Conclusions:

• OPT significantly improves generalizability and interpretability in multi-agent reinforcement learning.
• The proposed approach offers a robust solution for handling complex entity interactions.