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 > Artificial Intelligence

Title: NQMIX: Non-monotonic Value Function Factorization for Deep Multi-Agent Reinforcement Learning

Authors: Quanlin Chen
Abstract: Multi-agent value-based approaches recently make great progress, especially value decomposition methods. However, there are still a lot of limitations in value function factorization. In VDN, the joint action-value function is the sum of per-agent action-value function while the joint action-value function of QMIX is the monotonic mixing of per-agent action-value function. To some extent, QTRAN reduces the limitation of joint action-value functions that can be represented, but it has unsatisfied performance in complex tasks. In this paper, in order to extend the class of joint value functions that can be represented, we propose a novel actor-critic method called NQMIX. NQMIX introduces an off-policy policy gradient on QMIX and modify its network architecture, which can remove the monotonicity constraint of QMIX and implement a non-monotonic value function factorization for the joint action-value function. In addition, NQMIX takes the state-value as the learning target, which overcomes the problem in QMIX that the learning target is overestimated. Furthermore, NQMIX can be extended to continuous action space settings by introducing deterministic policy gradient on itself. Finally, we evaluate our actor-critic methods on SMAC domain, and show that it has a stronger performance than COMA and QMIX on complex maps with heterogeneous agent types. In addition, our ablation results show that our modification of mixer is effective.
Subjects: Artificial Intelligence (cs.AI); Multiagent Systems (cs.MA)
Cite as: arXiv:2104.01939 [cs.AI]
  (or arXiv:2104.01939v4 [cs.AI] for this version)

Submission history

From: Quanlin Chen [view email]
[v1] Mon, 5 Apr 2021 14:56:09 GMT (455kb)
[v2] Sun, 18 Apr 2021 05:10:30 GMT (541kb)
[v3] Mon, 21 Jun 2021 14:29:54 GMT (709kb)
[v4] Tue, 13 Jul 2021 11:13:45 GMT (709kb)

Link back to: arXiv, form interface, contact.