{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/on-the-pitfalls-of-measuring-emergent","title":"On the Pitfalls of Measuring Emergent Communication","arxiv_id":"1903.05168","date":"2019-03-12","proceeding":null,"authors":["Ryan Lowe","Jakob Foerster","Y-Lan Boureau","Joelle Pineau","Yann Dauphin"],"abstract":"How do we know if communication is emerging in a multi-agent system? The vast\nmajority of recent papers on emergent communication show that adding a\ncommunication channel leads to an increase in reward or task success. This is a\nuseful indicator, but provides only a coarse measure of the agent's learned\ncommunication abilities. As we move towards more complex environments, it\nbecomes imperative to have a set of finer tools that allow qualitative and\nquantitative insights into the emergence of communication. This may be\nespecially useful to allow humans to monitor agents' behaviour, whether for\nfault detection, assessing performance, or even building trust. In this paper,\nwe examine a few intuitive existing metrics for measuring communication, and\nshow that they can be misleading. Specifically, by training deep reinforcement\nlearning agents to play simple matrix games augmented with a communication\nchannel, we find a scenario where agents appear to communicate (their messages\nprovide information about their subsequent action), and yet the messages do not\nimpact the environment or other agent in any way. We explain this phenomenon\nusing ablation studies and by visualizing the representations of the learned\npolicies. We also survey some commonly used metrics for measuring emergent\ncommunication, and provide recommendations as to when these metrics should be\nused.","url_abs":"http://arxiv.org/abs/1903.05168v1","url_pdf":"http://arxiv.org/pdf/1903.05168v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"on-the-pitfalls-of-measuring-emergent","repo_url":"https://github.com/facebookresearch/measuring-emergent-comm","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":0,"framework":"pytorch","reach":{"status":"ok","spdx":"NOASSERTION"}}],"tasks":[{"task_slug":"deep-reinforcement-learning","task_name":"Deep Reinforcement Learning"},{"task_slug":"fault-detection","task_name":"Fault Detection"},{"task_slug":"reinforcement-learning","task_name":"Reinforcement Learning"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":"https://syntology.ai/paper/1903.05168","atlas_url":"https://app.syntology.ai/?focus=1903.05168","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}