Research map / Graph learning / GNN architectures
Generalized Value Iteration Networks:Life Beyond Lattices
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 |
|---|---|---|---|
| Use the embedding-based kernel instead of the directional or spatial kernel | GVIN on 100-node irregular graphs · 100-node irregular graphs | prediction accuracy: 57.90% / 57.45% -> 58.90% | Helped |
| Use the embedding-based kernel instead of the directional or spatial kernel | GVIN on 100-node irregular graphs · 100-node irregular graphs | path difference: 0.082 / 0.082 -> 0.079 | Helped |
| Train with episodic Q-learning instead of TRPO with curriculum learning | VIN · 16×16 mazes | success rate: 82.50% -> 98.67% | Helped |
| Train with episodic Q-learning instead of standard deep Q-learning | GVIN with embedding-based kernel · Irregular graphs | success rate | Helped |
| Train on 10-node graphs using episodic Q-learning rather than imitation learning | GVIN with embedding-based kernel · 100-node irregular graphs | success rate: 6.73% -> 100% | Helped |
| Train on 10-node graphs using episodic Q-learning rather than imitation learning | GVIN with embedding-based kernel · 100-node irregular graphs | path difference: 0.041 -> 0.148 | Mixed |
| Make the directional kernel direction-unaware rather than direction-aware (ablation) | GVIN on 100-node irregular graphs · 100-node irregular graphs | success rate: 34.75% -> 65.30% | Helped |
| Make the spatial kernel direction-unaware rather than direction-aware (ablation) | GVIN on 100-node irregular graphs · 100-node irregular graphs | success rate: 96.56% -> 97.17% | Helped |
| Use a spatial kernel instead of a directional kernel (ablation) | GVIN on 100-node irregular graphs · 100-node irregular graphs | success rate: 65.30% -> 97.17% | Helped |
| Exclude edge weights from the input (ablation) | GVIN trained with imitation learning · Weighted irregular graphs | success rate: 3.00% -> 97.60% | Helped |
| Exclude edge weights from the input (ablation) | GVIN trained with reinforcement learning · Weighted irregular graphs | expected reward: 0.943 -> 0.940 | Hurt |
| Use action-value rather than state-value imitation learning (ablation) | GVIN on 16×16 mazes · 2D mazes (16×16) | success rate: 99.57% -> 99.91% | No clear effect |
Closest papers by meaning
- Generalizing CNNs to graphs with learnable neighborhood quantization (NeurIPS 2024 · similarity 0.87)
- Scaling Value Iteration Networks to 5000 Layers for Extreme Long-Term Planning (ICML 2025 · similarity 0.86)
- Highway Value Iteration Networks (ICML 2024 · similarity 0.85)
- Learning More Expressive General Policies for Classical Planning Domains (AAAI 2025 · similarity 0.85)
- State Encodings for GNN-Based Lifted Planners (AAAI 2025 · similarity 0.85)
- Expressive Power of Graph Neural Networks for (Mixed-Integer) Quadratic Programs (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.