Research map / Reinforcement learning / MDPs and planning
BRIDGE: Bi-level Reinforcement Learning for Dynamic Group Structure in Coalition Formation Games
Shuqing Shi, Nam Phuong Tran, Hao Liang, Debmalya Mandal, Long Tran-Thanh, Yali Du
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 |
|---|---|---|---|
| reuse follower value network across coalition structures (vs reinitializing) (ablation) | BRIDGE · LBF 10a5t | baseline gain: 60.9 -> 89.3 | Improved |
| reinitialize follower policy network each outer iteration (vs reusing it) (ablation) | BRIDGE · LBF 10a5t | baseline gain: 81.9 -> 89.3 | Improved |
| increase follower training from 5 to 10 epochs (ablation) | BRIDGE leader · LBF 10a5t | baseline gain: 29.97 -> 89.29 | Improved |
| increase follower training from 7 to 10 epochs (ablation) | BRIDGE leader · LBF 10a5t | baseline gain: 88.35 -> 89.29 | Improved |
| increase follower training from 10 to 15 epochs (ablation) | BRIDGE leader · LBF 10a5t | baseline gain: 89.29 -> 91.28 | No clear effect |
| lengthen leader horizon from N/2 to N-1 (ablation) | BRIDGE · Modified normal, 100 agents | coalition value: 1187.4 -> 1596.2 | Improved |
| lengthen leader horizon from N-10 to N-1 (ablation) | BRIDGE · Modified normal, 100 agents | coalition value: 1344.7 -> 1596.2 | Improved |
| widen network from (64,64) to (128,128), 3 layers (ablation) | BRIDGE leader · Modified normal, 100 agents | coalition value: 1501.42 -> 1596.20 | Improved |
| deepen network from 3 to 5 layers (128,128) (ablation) | BRIDGE leader · Modified normal, 100 agents | coalition value: 1596.20 -> 1492.39 | No clear effect |
| continue training on target size 200 episodes instead of 100 (ablation) | BRIDGE pretrained on 3 agents · Modified normal, 5 agents | % of optimal value: 88.83% -> 97.2% | Improved |
| continue training on target size 200 episodes instead of 100 (ablation) | BRIDGE pretrained on 3 agents · Modified normal, 10 agents | % of optimal value: 41.61% -> 48.24% | Improved |
| pretrain on 3 agents plus few-shot (BRIDGE) vs training baseline CSG-UCT from scratch on same budget | BRIDGE · Modified normal, 10 agents | % of optimal value: 40.81% -> 48.24% | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- GRASP — GRASP: Navigating Retrosynthetic Planning with Goal-driven Policy (NeurIPS 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Deviate or Not: Learning Coalition Structures with Multiple-bit Observations in Games (AAAI 2025 · similarity 0.88)
- Online Learning of Group-Stable Coalition Structures (NeurIPS 2026 · similarity 0.87)
- Learning Coalition Structures with Games (AAAI 2024 · similarity 0.87)
- Online Learning of Coalition Structures by Selfish Agents (AAAI 2025 · similarity 0.87)
- Group-Graph Policy Optimization for Long-Horizon Agentic Reinforcement Learning (NeurIPS 2026 · similarity 0.84)
- HYGMA: Hypergraph Coordination Networks with Dynamic Grouping for Multi-Agent Reinforcement Learning (ICML 2025 · similarity 0.84)
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.
