Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Learning Multiagent Communication with Backpropagation
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 |
|---|---|---|---|
| add continuous broadcast communication between agents (CommNet) vs independent controllers | lever pulling game · Lever pulling game | fraction of distinct levers pulled: 0.59 -> 0.99 (supervised), 0.59 -> 0.94 (RL) | Helped |
| add continuous communication (CommNet) vs no communication | traffic junction (hard, MLP) · Traffic junction hard | failure rate (%): 20.6 -> 2.2 | Helped |
| add continuous communication with LSTM module | traffic junction (hard) · Traffic junction hard | failure rate (%): 9.4 -> 1.6 | Helped |
| fully-connected instead of modular communication | traffic junction and combat · Traffic junction hard; Combat | failure rate (%) / win rate (%): traffic MLP 20.6->12.5 (better); combat MLP 34.2->17.7 (worse) | Mixed |
| discrete symbol communication via RL vs no communication | traffic junction and combat · Traffic easy; combat LSTM | failure rate (%) / win rate (%): easy MLP 15.8->1.1 (better); hard MLP 20.6->15.8 (better); combat MLP 34.2->29.1 (worse); combat LSTM 44.3->46.4 (better | Mixed |
| add continuous communication (CommNet) vs no communication | combat task (MLP, RNN, LSTM) · Combat task | win rate (%): MLP 34.2->44.5; RNN 37.3->44.4; LSTM 44.3->49.5 | Helped |
| add continuous communication vs no communication at larger team size and better vision | combat task variants · Combat m=3, m=10, 5x5 vision | win rate (%): m=3 29.2->51.0; m=10 30.5->45.4; 5x5 vision 60.5->73.0 | Helped |
| continuous communication (CommNet MLP) vs independent MLP on bAbI | bAbI 10K · bAbI 10K | mean error (%): 15.2 -> 7.1 | Helped |
| add local connectivity communication instead of broadcast | traffic junction hard (RNN) · Traffic junction hard 4-junction | failure rate (%): 22.5 -> 21.1 | Helped |
| reduce agent visibility to zero (communication matters more) | traffic junction · Traffic junction | success rate with zero visibility: CommNet succeeds 90% of the time with zero visibility | No clear effect |
Closest papers by meaning
- Learning Multi-Agent Communication from Graph Modeling Perspective (ICLR 2024 · similarity 0.88)
- Language Grounded Multi-agent Reinforcement Learning with Human-interpretable Communication (NeurIPS 2024 · similarity 0.87)
- G-Designer: Architecting Multi-agent Communication Topologies via Graph Neural Networks (ICML 2025 · similarity 0.87)
- Learning Multi-Agent Communication with Contrastive Learning (ICLR 2024 · similarity 0.86)
- Expressive Multi-Agent Communication via Identity-Aware Learning (AAAI 2024 · similarity 0.86)
- Autonomous Agents for Collaborative Task under Information Asymmetry (NeurIPS 2024 · similarity 0.86)
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.