Research map / Multi-agent systems and games / Games and equilibria
Joint Policy Search for Multi-agent Collaboration with Imperfect Information
Yuandong Tian, Qucheng Gong, Yu Jiang
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 tabular JPS to CFR1k | CFR1k policies · Simple collaborative games | average reward: CFR1k 0.85 -> CFR1k+JPS 1.00 (Comm L=5); CFR1k 2.60 -> CFR1k+JPS 2.74 (2-Suit Mini-Bridge N=5) | Improved |
| Jointly search the current player's and partner's policies | A2C baseline · Contract Bridge bidding vs 10 independent A2C baselines | IMPs/b: 0.27 $\pm$ 0.13 -> 0.66 $\pm$ 0.11 | Improved |
| Increase the fraction of games using JPS from 5% to 10% (ablation) | JPSBid · Contract Bridge bidding vs 10 independent A2C baselines | IMPs/b: 0.66 $\pm$ 0.11 -> 0.39 $\pm$ 0.11 | Hurt |
| Use single-player search instead of joint policy search (ablation) | A2C baseline · Contract Bridge bidding vs baseline | IMPs/b: 0.20 -> 0.46 | Improved |
| Add explicit belief-modeling auxiliary loss (ablation) | A2C baseline · Contract Bridge bidding | IMPs: 2.99 $\pm$ 0.04 -> 2.53 $\pm$ 0.27 | Hurt |
| Increase network blocks from 2 to 4 (ablation) | A2C baseline · Contract Bridge bidding | IMPs: 2.97 $\pm$ 0.05 -> 2.99 $\pm$ 0.04 | No clear effect |
| Increase sync frequency from 1 to 50 (ablation) | A2C baseline · Contract Bridge bidding | IMPs: 2.89 $\pm$ 0.13 -> 2.99 $\pm$ 0.04 | No clear effect |
| Add random simultaneous exploration by the agent and partner (ablation) | A2C baseline · Contract Bridge bidding | IMPs: 2.99 $\pm$ 0.04 -> 2.37 $\pm$ 0.31 | Hurt |
| Use 1 sampled state per infoset rather than all states (ablation) | CFR1k+JPS · Mini-Hanabi | average reward: 9.50 -> 10.00 | Improved |
| Increase sampled states per infoset from 1 to 15 (ablation) | CFR1k+JPS · Mini-Hanabi | average reward: 10.00 -> 9.51 | Hurt |
| Use sample-based JPS rather than all-state JPS (ablation) | CFR1k initialization · Simple Bidding N=16 | average reward: 10.56 -> 10.61 | Improved |
| Use sample-based JPS rather than all-state JPS (ablation) | BAD initialization · Simple Bidding N=16 | average reward: 10.47 -> 9.91 (1 sample); 10.47 -> 10.57 (25 samples) | Mixed |
Closest papers by meaning
- Training Generalizable Collaborative Agents via Strategic Risk Aversion (NeurIPS 2026 · similarity 0.89)
- A Policy-Gradient Approach to Solving Imperfect-Information Games with Best-Iterate Convergence (ICLR 2025 · similarity 0.89)
- Incentivize without Bonus: Provably Efficient Model-based Online Multi-agent RL for Markov Games (ICML 2025 · similarity 0.88)
- Aligning Individual and Collective Objectives in Multi-Agent Cooperation (NeurIPS 2024 · similarity 0.88)
- Deviate or Not: Learning Coalition Structures with Multiple-bit Observations in Games (AAAI 2025 · similarity 0.87)
- Optimistic Policy Gradient in Multi-Player Markov Games with a Single Controller: Convergence beyond the Minty Property (AAAI 2024 · 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.
