Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Learning to Ground Multi-Agent Communication with Autoencoders
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 autoencoding of observations to ground communication messages | decentralized multi-agent communication · CIFAR Game | average reward: 0.142 ± 0.019 -> 0.348 ± 0.041 | Helped |
| Use autoencoding of observations to ground communication messages | decentralized multi-agent communication · RedBlueDoors | average reward: 0.729 ± 0.072 -> 0.984 ± 0.002 | Helped |
| Use autoencoding of observations to ground communication messages | decentralized multi-agent communication · FindGoal | average episode length: 158.0 ± 12.3 -> 103.5 ± 20.2 | Helped |
| Add a policy trained with reinforcement learning on top of the autoencoded representation (ablation) | ae-comm · CIFAR Game | average reward | Hurt |
| Add a policy trained with reinforcement learning on top of the autoencoded representation (ablation) | ae-comm · RedBlueDoors | average reward | Hurt |
| Add a policy trained with reinforcement learning on top of the autoencoded representation (ablation) | ae-comm · FindGoal | average episode length | No clear effect |
| Optimize the autoencoder and RL objectives alternately, freezing one component while training the other (ablation) | ae-comm · communication training experiments | task performance | No clear effect |
| Continue policy-optimization training for up to 1 million iterations (ablation) | communication trained with autoencoding and RL · communication training experiments | task performance: 1 million | No clear effect |
| Use autoencoded messages instead of RL-trained communication | communication with a discrete channel · CIFAR Game | average reward: 0.099 ± 0.013 -> 0.348 ± 0.041 | Helped |
| Use autoencoded messages instead of RL-trained communication | communication with a discrete channel · RedBlueDoors | average reward: 0.174 ± 0.009 -> 0.984 ± 0.002 | Helped |
| Use autoencoded messages instead of policy-derived discretized states | communication with a fixed-size discrete channel · CIFAR Game | average reward: 0.107 ± 0.021 -> 0.348 ± 0.041 | Helped |
| Use autoencoded messages instead of policy-derived discretized states | communication with a fixed-size discrete channel · RedBlueDoors | average reward: 0.382 ± 0.022 -> 0.984 ± 0.002 | Helped |
Closest papers by meaning
- Language Grounded Multi-agent Reinforcement Learning with Human-interpretable Communication (NeurIPS 2024 · similarity 0.91)
- Learning Multi-Agent Communication with Contrastive Learning (ICLR 2024 · similarity 0.90)
- Learning Multiagent Communication with Backpropagation (NeurIPS 2016 · similarity 0.89)
- Ground-Compose-Reinforce: Grounding Language in Agentic Behaviours using Limited Data (NeurIPS 2025 · similarity 0.87)
- Learning to Orchestrate Agents in Natural Language with the Conductor (ICLR 2026 · similarity 0.87)
- Building Cooperative Embodied Agents Modularly with Large Language Models (ICLR 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.