We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations

DBLP - CS Bibliography


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Multiagent Systems

Title: Learning Multi-agent Action Coordination via Electing First-move Agent

Abstract: Learning to coordinate actions among agents is essential in complicated multi-agent systems. Prior works are constrained mainly by the assumption that all agents act simultaneously, and asynchronous action coordination between agents is rarely considered. This paper introduces a bi-level multi-agent decision hierarchy for coordinated behavior planning. We propose a novel election mechanism in which we adopt a graph convolutional network to model the interaction among agents and elect a first-move agent for asynchronous guidance. We also propose a dynamically weighted mixing network to effectively reduce the misestimation of the value function during training. This work is the first to explicitly model the asynchronous multi-agent action coordination, and this explicitness enables to choose the optimal first-move agent. The results on Cooperative Navigation and Google Football demonstrate that the proposed algorithm can achieve superior performance in cooperative environments. Our code is available at \url{this https URL}.
Comments: Accepted by ICAPS 2022
Subjects: Multiagent Systems (cs.MA); Computer Science and Game Theory (cs.GT)
Cite as: arXiv:2110.08126 [cs.MA]
  (or arXiv:2110.08126v3 [cs.MA] for this version)

Submission history

From: Jingqing Ruan [view email]
[v1] Mon, 20 Sep 2021 05:25:32 GMT (7173kb,D)
[v2] Sat, 19 Feb 2022 11:37:29 GMT (8082kb,D)
[v3] Thu, 24 Feb 2022 14:49:00 GMT (8084kb,D)

Link back to: arXiv, form interface, contact.