Research map / Multi-agent systems and games / Cooperative multi-agent RL
I2Q: A Fully Decentralized Q-Learning Algorithm
Jiechuan Jiang, Zongqing Lu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 I2Q (learning on modeled ideal transition) instead of IQL | independent Q-learning · Random matrix games, L=3..6 | normalized reward | Improved |
| Increase exploration rate epsilon from 0.1 to 0.7 | I2Q · Matrix games L=4 | normalized reward / convergence speed | Improved |
| Increase exploration rate epsilon from 0.1 to 0.7 | IQL · Matrix games L=4 | normalized reward | Hurt |
| Use I2Q instead of IQL on one-stage QTRAN/QPLEX matrix games | independent Q-learning · Matrix game case 1 and case 2 | reward | Improved |
| Use I2Q instead of IDDPG, Hysteretic IDDPG and D3G | decentralized DDPG · MPE differential games N=2..5 | reward | Improved |
| Add optimistic (hysteretic) update to IDDPG | IDDPG · MPE differential games | reward | No clear effect |
| Apply D3G forward model independently per agent | IDDPG · MPE differential games | reward | No clear effect |
| Set lambda to 0 (no Qss maximization; reduces to MBPO) (ablation) | I2Q · MPE differential games | reward | Hurt |
| Set lambda too large (10) (ablation) | I2Q · MPE differential games | reward | Hurt |
| Use moderate lambda 0.5 (ablation) | I2Q · MPE differential games N=2..5 | reward | Improved |
| Add transition noise xi 0.02 and 0.03 (ablation) | I2Q · MPE differential games N=3 | reward | No clear effect |
| Add strong transition noise xi 0.04 and 0.05 (ablation) | I2Q · MPE differential games N=3 | reward | Hurt |
Closest papers by meaning
- Decentralized Q-learning in Zero-sum Markov Games (NeurIPS 2021 · similarity 0.90)
- Multi-Agent Determinantal Q-Learning (ICML 2020 · similarity 0.89)
- Consensus Learning for Cooperative Multi-Agent Reinforcement Learning (AAAI 2023 · similarity 0.89)
- F2A2: Flexible Fully-decentralized Approximate Actor-critic for Cooperative Multi-agent Reinforcement Learning (JMLR 2023 · similarity 0.88)
- Mean-Field Sampling for Cooperative Multi-Agent Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- Inverse Factorized Soft Q-Learning for Cooperative Multi-agent Imitation Learning (NeurIPS 2024 · similarity 0.87)
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.
