Research map / Reinforcement learning / Visual and representation RL
Graph Reinforcement Learning for Network Control via Bi-Level Optimization
Daniele Gammelli, James Harrison, Kaidi Yang, Marco Pavone, Filipe Rodrigues, Francisco C. Pereira
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 |
|---|---|---|---|
| Bi-level formulation (RL outputs desired next state, LCP computes flows) instead of end-to-end RL (ablation) | SCIM and DVR problems · SCIM 1F2S, 1F3S, 1F10S; DVR New York, Shenzhen | % of oracle profit: e.g. 1F2S 96.8% vs 95.8% (GNN end-to-end); DVR NY 94.9% vs 17.2% | Improved |
| Use MPNN (max aggregation) backbone instead of GCN/GAT/MLP within bi-level policy (ablation) | Minimum cost flow · Minimum cost flow, all scenarios | % of oracle performance: MPNN 89.7% vs GCN 31.3% on 2-hops; avg 86.7% vs others | Improved |
| Distance term to desired next state included in inner problem (vs greedy one-step argmin of reward only) (ablation) | SCIM and DVR · SCIM 1F2S/1F3S/1F10S; DVR NY/Shenzhen | reward / served demand: SCIM 1F2S reward -102,919 to 192; DVR NY served demand 1,357 to 1,824 | Improved |
| Bi-level MPNN policy vs S-type order-up-to heuristic | SCIM · SCIM 1F2S, 1F3S, 1F10S | % of oracle profit: 96.8% vs 75.5% (1F2S); 96.0% vs 84.2% (1F3S); 99.5% vs 86.4% (1F10S) | Improved |
| Bi-level graph-RL vs evenly-balanced domain heuristic for vehicle rebalancing | DVR · New York, Shenzhen | % of oracle profit: 94.9% vs 71.6% (NY); 83.8% vs 50.1% (Shenzhen) | Improved |
| GNN-based backbone vs MLP in bi-level policy, dynamic topology (ablation) | Minimum cost flow dynamic topology · Dynamic topology | % of oracle performance: MPNN 83.9% vs GCN 53.4%; MLP-RL not applicable (N/A) | Improved |
Closest papers by meaning
- Local Policies for Graph-Structured Markov Decision Processes (ICML 2026 · similarity 0.88)
- Learning Combinatorial Optimization Algorithms over Graphs (NeurIPS 2017 · similarity 0.88)
- GraphOpt: Learning Optimization Models of Graph Formation (ICML 2020 · similarity 0.87)
- Optimal Transport–Guided Stochastic Control for Graph Combinatorial Optimization (ICML 2026 · similarity 0.87)
- Bayesian Ego-graph inference for Networked Multi-Agent Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
- Bridging Training and Execution via Dynamic Directed Graph-Based Communication in Cooperative Multi-Agent Systems (AAAI 2025 · similarity 0.85)
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.
