AAC: Differentiable Landmark Compression for ALT Heuristics
The AAC (Architecturally Admissible Compressor) serves as a differentiable module for landmark selection in ALT (A*, Landmarks, and Triangle inequality) shortest-path heuristics. Its design guarantees admissibility, as each forward pass produces a row-stochastic mixture of lower bounds based on the triangle inequality, ensuring admissibility across all parameter configurations without requiring convergence, calibration, or projection. During deployment, the module simplifies to classical ALT on a learned subset, seamlessly integrating with neural encoders while maintaining the traditional toolchain. This represents the inaugural differentiable approach within the compress-while-preserving-admissibility framework in heuristic search. Furthermore, under a matched per-vertex memory protocol, FPS-ALT demonstrates near-optimal coverage on metric graphs, maintaining only a slight percentage of headroom.
Key facts
- AAC is a differentiable landmark-selection module for ALT heuristics.
- Outputs are admissible by construction via row-stochastic mixtures.
- No convergence, calibration, or projection needed.
- At deployment, reduces to classical ALT on a learned subset.
- Composes end-to-end with neural encoders.
- First differentiable instance of compress-while-preserving-admissibility.
- FPS-ALT has provably near-optimal coverage on metric graphs.
- Memory protocol leaves at most a few percentage points of headroom.
Entities
—