Research map / Multi-agent systems and games / Games and equilibria
Improving Policies via Search in Cooperative Partially Observable Games
Adam Lerer, Hengyuan Hu, Jakob Foerster, Noam Brown
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 |
|---|---|---|---|
| Add single-agent search to the agreed blueprint | Hanabi policies including SAD, SmartBot, CloneBot, and DQN · Hanabi 2-player | average score: 22.99 ± 0.001 -> 24.21 ± 0.002 (SmartBot); 24.08 ± 0.01 -> 24.53 ± 0.01 (SAD) | Improved |
| Add retrospective multi-agent search after single-agent search | SmartBot, DQN, and SAD blueprints · Hanabi 2-player | average score: 24.21 ± 0.002 -> 24.46 ± 0.01 (SmartBot); 24.53 ± 0.01 -> 24.61 ± 0.01 (SAD) | Improved |
| Add single-agent search to SAD | SAD · Hanabi 3-player | average score: 23.99 ± 0.01 -> 24.61 ± 0.01 | Improved |
| Add single-agent search to SAD | SAD · Hanabi 4-player | average score: 23.81 ± 0.01 -> 24.49 ± 0.01 | Improved |
| Add single-agent search to SAD | SAD · Hanabi 5-player | average score: 23.01 ± 0.01 -> 23.91 ± 0.01 | Improved |
| Increase the max range for retrospective multi-agent search (ablation) | SmartBot · Hanabi 2-player | average score: 24.21 ± 0.002 -> 24.48 ± 0.01 (max range 0 -> 10,000) | Improved |
| Increase the max range for retrospective multi-agent search (ablation) | SmartBot · Hanabi 2-player | rollouts: 1.5 × 10^5 -> 1.8 × 10^8 (max range 0 -> 10,000) | Hurt |
| Use independent multi-agent search instead of coordinated retrospective search (ablation) | SmartBot · Hanabi 2-player | average score: 22.99 -> 14.41 | Hurt |
| Raise the deviation threshold from 0 to 0.2 for independent search with belief uncertainty 0.05 (ablation) | SmartBot · Hanabi 2-player | average score: 14.41 -> 24.18 | Improved |
| Use a deviation threshold of 0.05 rather than allowing all estimated improvements (ablation) | SmartBot search · Hanabi | average score: null | Improved |
| Use UCB-like pruning to skip rollouts for clearly inferior actions (ablation) | SPARTA search · Hanabi | average score: null | No clear effect |
| Use UCB-like pruning and the blueprint deviation threshold (ablation) | SPARTA search · Hanabi | rollouts per timestep: 10× fewer rollouts | Improved |
Closest papers by meaning
- Robust and Diverse Multi-Agent Learning via Rational Policy Gradient (NeurIPS 2025 · similarity 0.88)
- Opponent Modeling based on Subgoal Inference (NeurIPS 2024 · similarity 0.88)
- Multi-agent cooperation through learning-aware policy gradients (ICLR 2025 · similarity 0.88)
- FightLadder: A Benchmark for Competitive Multi-Agent Reinforcement Learning (ICML 2024 · similarity 0.87)
- Mimicking To Dominate: Imitation Learning Strategies for Success in Multiagent Games (NeurIPS 2024 · similarity 0.87)
- MARSHAL: Incentivizing Multi-Agent Reasoning via Self-Play with Strategic LLMs (ICLR 2026 · 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.
