Reproduce “Learning to Ground Multi-Agent Communication with Autoencoders” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Toru Lin, Minyoung Huh, Chris Stauffer, Ser-Nam Lim, Phillip Isola
This paper presents a method for grounding multi-agent communication using autoencoders, enabling decentralized agents to develop effective communication protocols in cooperative tasks. The approach consistently outperformed baselines across various environments, facilitating improved coordination and task performance by establishing a shared understanding of environmental cues.
Key insights
- Formulating communication grounding as a representation learning problem using autoencoders allows agents to develop a shared understanding of their environment.
- Training a policy jointly with the autoencoder representation can degrade performance due to a lack of correlation between visual observation and communication rewards.
- Agents learn to cluster their communications based on meaningful game states, indicating the development of a grounded visual language.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.