Research map / Multi-agent systems and games / Cooperative multi-agent RL
Selectively Sharing Experiences Improves Multi-Agent Reinforcement Learning
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 selective sharing of high-td-error experiences | DDQN · Pursuit | total mean episode reward: 181.43 (+-4.16) -> 454.56 (+-5.89) | Helped |
| Add selective sharing of high-td-error experiences | DDQN · Battle | blue team total mean episode reward: 9.46 (+-0.86) -> 17.05 (+-2.74) | Helped |
| Add selective sharing of high-td-error experiences | DDQN · Adversarial-Pursuit | prey team total mean episode reward: -719.78 (+-66.82) -> -506.29 (+-35.03) | Helped |
| Share all experiences indiscriminately instead of selecting a small relevant subset (ablation) | SUPER-DDQN · Pursuit | total mean episode reward: 454.56 (+-5.89) -> 148.27 (+-9.22) | Hurt |
| Share all experiences indiscriminately instead of selecting a small relevant subset (ablation) | SUPER-DDQN · Battle | blue team total mean episode reward: 17.05 (+-2.74) -> 9.25 (+-6.60) | Hurt |
| Share all experiences indiscriminately instead of selecting a small relevant subset (ablation) | SUPER-DDQN · Adversarial-Pursuit | prey team total mean episode reward: -506.29 (+-35.03) -> -568.91 (+-40.82) | Hurt |
| Select a small fraction of experiences uniformly at random rather than by td-error priority (ablation) | SUPER-DDQN · Pursuit and Battle | performance | Hurt |
| Use stochastic td-error-weighted selection instead of deterministic quantile selection (ablation) | SUPER-DDQN · Pursuit and Battle | total mean episode reward: 454.56 (+-5.89) -> 266.42 (+-85.52); 17.05 (+-2.74) -> 7.43 (+-5.64) | Hurt |
| Use Gaussian mean-and-variance selection instead of quantile selection (ablation) | SUPER-DDQN · Pursuit and Battle | total mean episode reward: 454.56 (+-5.89) -> 373.63 (+-9.15); 17.05 (+-2.74) -> 16.28 (+-3.42) | Mixed |
| Share experiences at target bandwidth around 0.01-0.1 instead of sharing none or all (ablation) | SUPER-DDQN quantile · Pursuit | converged performance: 0.01 - 0.1 | Helped |
| Share selected experiences with other agents | DQN and dueling DDQN · Multiple benchmark environments and hyperparameter settings | performance | Helped |
| Use parameter sharing instead of selective experience sharing | DDQN · Adversarial-Pursuit | prey team total mean episode reward: -506.29 (+-35.03) -> -268.78 (+-60.11) | Hurt |
Closest papers by meaning
- HyperMARL: Adaptive Hypernetworks for Multi-Agent RL (NeurIPS 2025 · similarity 0.89)
- DIFFER:Decomposing Individual Reward for Fair Experience Replay in Multi-Agent Reinforcement Learning (NeurIPS 2023 · similarity 0.89)
- Cautiously-Optimistic Knowledge Sharing for Cooperative Multi-Agent Reinforcement Learning (AAAI 2024 · similarity 0.89)
- Multi-agent cooperation through learning-aware policy gradients (ICLR 2025 · similarity 0.87)
- Learning to Share: Selective Memory for Efficient Parallel Agentic Systems (ICML 2026 · similarity 0.86)
- A Generalist Hanabi Agent (ICLR 2025 · 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.