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 > Machine Learning

Title: Towards Understanding Linear Value Decomposition in Cooperative Multi-Agent Q-Learning

Abstract: Value decomposition is a popular and promising approach to scaling up multi-agent reinforcement learning in cooperative settings. However, the theoretical understanding of such methods is limited. In this paper, we introduce a variant of the fitted Q-iteration framework for analyzing multi-agent Q-learning with value decomposition. Based on this framework, we derive a closed-form solution to the Bellman error minimization with linear value decomposition. With this novel solution, we further reveal two interesting insights: 1) linear value decomposition implicitly implements a classical multi-agent credit assignment called counterfactual difference rewards; and 2) multi-agent Q-learning with linear value decomposition requires on-policy data distribution to achieve numerical stability. In the empirical study, our experiments demonstrate the realizability of our theoretical implications in a broad set of complicated tasks. They show that most state-of-the-art deep multi-agent Q-learning algorithms using linear value decomposition cannot efficiently utilize off-policy samples, which may even lead to an unbounded divergence.
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Multiagent Systems (cs.MA); Machine Learning (stat.ML)
Cite as: arXiv:2006.00587 [cs.LG]
  (or arXiv:2006.00587v2 [cs.LG] for this version)

Submission history

From: Zhizhou Ren [view email]
[v1] Sun, 31 May 2020 19:14:03 GMT (290kb,D)
[v2] Tue, 23 Jun 2020 15:11:41 GMT (386kb,D)
[v3] Mon, 5 Oct 2020 15:24:13 GMT (1160kb,D)
[v4] Wed, 29 Sep 2021 11:07:05 GMT (1090kb,D)
[v5] Sun, 31 Oct 2021 06:21:48 GMT (1128kb,D)

Link back to: arXiv, form interface, contact.