Emergent Robust Communication for Multi-Round Interactions in Noisy Environments

Fábio Vital, Alberto Sardinha, Francisco S. Melo

Primary Area: reinforcement learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: emergent communication, reinforcement learning, multi-agent reinforcement learning, transfer learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: This work explores new neural-agent architectures that can develop general and robust communication protocols for environment like the Lewis Game and more complex extensions.

Abstract: We contribute a novel multi-agent architecture capable of learning a discrete communication protocol without any prior knowledge of the task to solve. We focus on ensuring agents can create a common language during their training to be able to cooperate and solve the task at hand, which is one of the primary goals of the emergent communication field. On top of this, we focus on increasing the task's difficulty by creating a novel referential game, based on the original Lewis Game, that has two new sources of complexity: adding random noise to the message being transmitted and the capability for multiple interactions between the agents before making a final prediction. When evaluating the proposed architecture on the newly developed game, we observe that the emerging communication protocol's generalization aptitude remains equivalent to architectures employed in much simpler and elementary games. Additionally, our method is the only one suitable to produce robust communication protocols that can handle cases with and without noise while maintaining increased generalization performance levels.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2673