Silence-Conditioned Posterior Estimation for Sparse Multi-Agent Communication over Lossy Channels
A missing packet in cooperative multi-agent systems has two causes: the sender stayed silent, or the channel erased the packet. Conflating them discards the information in intentional silence and makes receivers over-confident as losses grow. We formalise the observation model, derive the absence likelihood separating the two causes, and characterise when silence is informative. Silence-Conditioned Posterior Estimation, or SCOPE, learns a message prior and a token-conditioned model of the sender’s gate and, given a calibrated erasure probability, combines them at execution time; it trains with a gated recurrent multi-agent proximal policy optimisation actor-critic. Across an analytic study, an offline study and two cooperative tasks, the mixture is invariant to the loss rate and reports a calibrated weight for the cause of an absence. The benefit is conditional on the channel being worth reading. In blind navigation, where only a scout sees the target, packets are worth 4.9 return on a clean channel and 1.9 at 70% loss, and at 60% loss SCOPE attains the best absence log score of any packet-reading receiver. In foraging, food is visible to every agent, so the channel is redundant and the receivers differ in belief quality rather than in return.
Authors
- Zongzhe Nie
- Jinyin Bai (ORCID: https://orcid.org/0009-0007-6677-0971)
- Shiluo Guo
- KaiYang Kou
- Tianjin Ni
- Xiangchen Wang (ORCID: https://orcid.org/0009-0001-8283-1734)
- Jinji Zhou (ORCID: https://orcid.org/0009-0005-4948-4610)
- Mingxi Wang (ORCID: https://orcid.org/0009-0006-8986-5896)
- Wei Zhu (ORCID: https://orcid.org/0009-0008-5777-9621)
- Hang Liu
- Chengjie Lu
- Yanlin Wu
- Rui Tang
- Bo Liu
- Yongjun Peng
Institutions
- National University of Defense Technology (CN)
- Beijing Satellite Navigation Center (CN)
Publication Details
- Journal
- Electronics
- Published
- 2026-11-16
- DOI
- https://doi.org/10.3390/electronics15225332
- Primary Topic
- Reinforcement Learning in Robotics
- Type
- article
- Field-Weighted Citation Impact
- 0.00