2021
DOI: 10.48550/arxiv.2102.03479
|View full text |Cite
Preprint
|
Sign up to set email alerts
|

Rethinking the Implementation Tricks and Monotonicity Constraint in Cooperative Multi-Agent Reinforcement Learning

Help me understand this report

Search citation statements

Order By: Relevance

Paper Sections

Select...
3
1
1

Citation Types

0
19
0

Year Published

2022
2022
2024
2024

Publication Types

Select...
4
4

Relationship

0
8

Authors

Journals

citations
Cited by 14 publications
(19 citation statements)
references
References 13 publications
0
19
0
Order By: Relevance
“…In this section, we show our experimental results compared to several state-of-theart algorithms, not limited to only multi-agent policy gradient methods, but also with decomposed value methods and combined methods on decentralized StarCraft II micromanagement benchmark [33], namely COMA [4], MAVEN [22], QMIX [32], and VDAC-vmix from VDAC [36] as they report out of the two proposed methods it deliv- We then perform several ablation studies to analyze the factors that contribute to the performance. It is worth noting that the StarCraft Multi-Agent Challenges (SMAC) are significantly affected by various code-level optimizations, i.e., hyper-parameter tuning, as also found by [13], some works are relying on heavy hyper-parameters tuning to achieve results that they otherwise cannot. Consistent with previous work, we carry out the test with the same hyper-parameters settings across all algorithms.…”
Section: Decentralised Starcraft II Micromanagement Benchmarkmentioning
confidence: 99%
“…In this section, we show our experimental results compared to several state-of-theart algorithms, not limited to only multi-agent policy gradient methods, but also with decomposed value methods and combined methods on decentralized StarCraft II micromanagement benchmark [33], namely COMA [4], MAVEN [22], QMIX [32], and VDAC-vmix from VDAC [36] as they report out of the two proposed methods it deliv- We then perform several ablation studies to analyze the factors that contribute to the performance. It is worth noting that the StarCraft Multi-Agent Challenges (SMAC) are significantly affected by various code-level optimizations, i.e., hyper-parameter tuning, as also found by [13], some works are relying on heavy hyper-parameters tuning to achieve results that they otherwise cannot. Consistent with previous work, we carry out the test with the same hyper-parameters settings across all algorithms.…”
Section: Decentralised Starcraft II Micromanagement Benchmarkmentioning
confidence: 99%
“…Almost all scenarios contain homogeneous agents, thus SMAC is suitable for agent permutation invariant study. For the reason that the Easy scenarios can be effectively solved by existing methods (Hu et al, 2021a), we mainly evaluate our methods in all Hard and Super Hard scenarios.…”
Section: Experimental Setupsmentioning
confidence: 99%
“…Evaluation Metric. Following Rashid et al (2018); Samvelyan et al (2019); Hu et al (2021a), our evaluation metric is the function that maps the environment steps to winning percentage throughout training. For each experiment, we run 32 test episodes without exploration to record the test win rate.…”
Section: Experimental Setupsmentioning
confidence: 99%
See 1 more Smart Citation
“…Given a shared reward signal in the fully-cooperative setting, the agents could estimate the impact of the joint action on the common reward but lack the ability to estimate the impact of a single agent's action. Even in the centralized training scheme in which agents observe the global state and access to the joint action, the credit assignment problem is difficult to alleviate without strong assumptions on the reward structure [Hu et al, 2021;. Further, the partially observable setting and the decentralized training scheme bring in additional challenges to the credit assignment problem.…”
Section: Challenges In Marlmentioning
confidence: 99%