Research map / Multi-agent systems and games / Cooperative multi-agent RL
Benchmarking Multi-Agent Deep Reinforcement Learning Algorithms in Cooperative Tasks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use centralised state-value critic (MAA2C) instead of independent critic (IA2C) | A2C · RWARE Tiny 4p | maximum return: 26.34(460) -> 32.50(979) | Helped |
| use centralised state-value critic (MAA2C) instead of independent critic (IA2C) | A2C · LBF 15x15-3p-5f | maximum return: 0.89(4) -> 0.87(6)* | No clear effect |
| use centralised critic (MAPPO) instead of independent critic (IPPO) | PPO · RWARE Small 4p | maximum return: 19.78(312) -> 27.00(180) | Helped |
| use PPO surrogate with multiple epochs (MAPPO) instead of A2C (MAA2C) | centralised critic · RWARE Small 4p | maximum return: 10.30(148) -> 27.00(180) | Helped |
| use PPO instead of A2C (MAPPO vs MAA2C) | centralised critic · LBF 15x15-3p-5f | maximum return: 0.87(6)* -> 0.77(2) | No clear effect |
| use QMIX mixing network instead of linear VDN decomposition | value decomposition · MPE Tag | maximum return: 24.50(219) -> 31.18(381) | Helped |
| use QMIX mixing network instead of linear VDN decomposition | value decomposition · LBF | maximum return | No clear effect |
| use value decomposition (VDN/QMIX) instead of IQL | IQL · SMAC MMM2 | maximum return: 13.69(102) -> 18.40(24)* | Helped |
| use value decomposition (VDN/QMIX) in sparse-reward tasks | IQL · RWARE Tiny 4p | maximum return: 0.72(37) -> 0.30(19) | No clear effect |
| train with parameter sharing instead of separate parameters | MARL algorithms · all non-matrix environments | normalised maximum return | Helped |
| train with parameter sharing instead of separate parameters | MARL algorithms · matrix games | normalised maximum return | No clear effect |
| average over all algorithms: sharing difference significance | MARL algorithms · averages | normalised maximum return | Mixed |
Closest papers by meaning
- Revisiting Some Common Practices in Cooperative Multi-Agent Reinforcement Learning (ICML 2022 · similarity 0.94)
- Towards a Standardised Performance Evaluation Protocol for Cooperative MARL (NeurIPS 2022 · similarity 0.92)
- Value-Based Deep Multi-Agent Reinforcement Learning with Dynamic Sparse Training (NeurIPS 2024 · similarity 0.91)
- Enhancing Cooperative Multi-Agent Reinforcement Learning with State Modelling and Adversarial Exploration (ICML 2025 · similarity 0.90)
- HyperMARL: Adaptive Hypernetworks for Multi-Agent RL (NeurIPS 2025 · similarity 0.90)
- Dispelling the Mirage of Progress in Offline MARL through Standardised Baselines and Evaluation (NeurIPS 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.