Research map / Multi-agent systems and games / Cooperative multi-agent RL
Message-Dropout: An Efficient Training Method for Multi-Agent Deep Reinforcement Learning
Woojun Kim, Myungsik Cho, Youngchul Sung
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 |
|---|---|---|---|
| Apply message-dropout with a proper dropout rate | DCC in pursuit · Pursuit (N=6, N=8) | evader catches | Improved |
| Apply message-dropout with dropout rate 0.2 or 0.5 | DCC · Cooperative navigation (N=8, L=8; N=10, L=10) | sum of agents' rewards | Improved |
| Apply message-dropout with dropout rate 0.2 or 0.5 | MADDPG · Waterworld (N=8, K=4; N=10, K=5) | food target catches | Improved |
| Use message-dropout during training | DCC · Pursuit with half of communication links broken | evader catches | Improved |
| Use message-dropout during training | DCC · Pursuit with all communication links broken | evader catches | Hurt |
| Increase dropout rate (ablation) | DCC-MD · Pursuit training | learning speed | Improved |
| Use dropout rates 0.2 to 0.5 rather than other tested rates (ablation) | DCC-MD · Pursuit execution | evader catches | No clear effect |
| Use block-wise message-dropout instead of element-wise standard dropout (ablation) | DCC · Pursuit | performance | Improved |
| Keep each agent's own observation undropped (ablation) | DCC with message-dropout · Pursuit | training time | Improved |
| Apply dropout to each agent's own observation as well as other agents' messages (ablation) | DCC · Pursuit | training time | Hurt |
| Use the observation/message-separated Q-function architecture instead of concatenated inputs (ablation) | DCC-MD · Pursuit (N=8) and Waterworld | learning success | Improved |
| Compress messages with an autoencoder (ablation) | DCC and DCC-MD · Pursuit (N=8) | steady-state performance | Hurt |
Closest papers by meaning
- Learning Agent Communication under Limited Bandwidth by Message Pruning (AAAI 2020 · similarity 0.87)
- Multi-Agent Reinforcement Learning with Communication-Constrained Priors (NeurIPS 2025 · similarity 0.85)
- Learning to Schedule Communication in Multi-agent Reinforcement Learning (ICLR 2019 · similarity 0.85)
- RGMComm: Return Gap Minimization via Discrete Communications in Multi-Agent Reinforcement Learning (AAAI 2024 · similarity 0.84)
- Value-Based Deep Multi-Agent Reinforcement Learning with Dynamic Sparse Training (NeurIPS 2024 · similarity 0.83)
- Efficient Communication in Multi-Agent Reinforcement Learning with Implicit Consensus Generation (AAAI 2025 · similarity 0.83)
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.
