Research map / Architectures and efficiency / RNNs and normalization
Highway Value Iteration Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Embeds highway value iteration into value iteration networks (adding aggregate gate, value exploration modules, and filter gate) so VINs can be trained with hundreds of layers for long-horizon 2D/3D maze planning.
Problem
Value iteration networks struggle with long-term planning because very deep VINs are hard to train and their success rates collapse on long shortest-path tasks.
Approach
Replace the VIN planning module with a highway value iteration structure: parallel value-exploration modules using stochastic embedded policies (expectation over actions), an aggregate gate forming skip connections across layers, and a max-based filter gate that discards activations lower than the current value, preserving convergence guarantees.
Key results
- On 25x25 mazes with shortest path length 130-230, highway VIN at depth 300 reaches 90.06% success versus under 3% for the best VIN/GPPN depths shown.
- On 15x15 mazes with shortest path length 60-100, highway VIN at depth 100 reaches 98.61% success, compared with 76.16% for the best VIN (depth 40) and 65.17% for GPPN.
- Removing the filter gate substantially degrades highway VIN performance, and VIN's learned latent action entropy stays near zero at depth 200-300 whereas highway VIN's stays above 4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add highway value iteration components to VIN | VIN · 25×25 maze, SPL [130,230] | success rate: 2.96 \pm 0.66 -> 90.06 \pm 3.13 | Helped |
| Increase Highway VIN depth from N=60 to N=300 | Highway VIN · 25×25 maze, SPL [130,230] | success rate: +69.38% | Helped |
| Increase depth of a highway network | Highway network · 2D maze long-term planning | success rate | No clear effect |
| Increase depth of GPPN | GPPN · 25×25 maze, SPL [130,230] | success rate | No clear effect |
| Remove the filter gate (ablation) | Highway VIN · 2D maze navigation | success rate | Hurt |
| Add the filter gate (ablation) | Highway VIN · 2D maze navigation | success rate | Helped |
| Replace VE modules with VI modules (ablation) | Highway VIN · 25×25 maze navigation | success rate | Hurt |
| Add VE modules (ablation) | Highway VIN · 25×25 maze navigation | success rate | Helped |
| Increase parallel VE modules to N_p=3 rather than N_p=1 (ablation) | Highway VIN · 2D maze navigation, depth N=100 | success rate | Helped |
| Increase parallel VE modules rather than use N_p=1 (ablation) | Highway VIN · 2D maze navigation, depth N=300 | success rate | Hurt |
| Add parallel VE modules (ablation) | Highway VIN · 2D maze navigation across depths | success rate | Mixed |
Surprising findings
- Plain VIN success rates drop to nearly 0% for depths above 150, so deeper VINs perform worse rather than better.
- Highway networks with skip connections do not improve long-term planning despite stable training at depth, suggesting skip connections alone do not supply the needed planning inductive bias.
Limitations
- Evaluated only on small synthetic grid-maze and ViZDoom navigation tasks, not large-scale real-world planning.
- Ablation shows multiple parallel VE modules can hurt very deep networks, so the best configuration depends on depth.
- Highway VIN uses about 5x more GPU memory than VIN at 300 layers.
Experimental setup
Datasets: GPPN 2D maze navigation dataset (25K train, 5K val, 5K test) 3D ViZDoom maze navigation
Benchmarks: 2D maze navigation 15x15 2D maze navigation 25x25 3D ViZDoom maze navigation
Metrics: success rate optimality rate entropy of selected latent actions
Baselines: VIN GPPN Highway network
Builds on: value iteration networks highway value iteration highway networks gated path-planning networks dropout
Closest papers by meaning
- Scaling Value Iteration Networks to 5000 Layers for Extreme Long-Term Planning (ICML 2025 · similarity 0.91)
- Value Iteration Networks (NeurIPS 2016 · similarity 0.91)
- Gated Path Planning Networks (ICML 2018 · similarity 0.88)
- Information-based Value Iteration Networks for Decision Making Under Uncertainty (ICLR 2026 · similarity 0.86)
- Accelerated training through iterative gradient propagation along the residual path (ICLR 2025 · similarity 0.82)
- DPN: Decoupling Partition and Navigation for Neural Solvers of Min-max Vehicle Routing Problems (ICML 2024 · similarity 0.82)
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.