Show HN: Solving Sudoku reasoning via Energy Geometric models
13 days ago
- #Constraint Satisfaction
- #GPU-acceleration
- #Riemannian Geometry
- A GPU-accelerated solver achieves a 1,226× speedup over Python CPU for solving Sudoku and generalizes to any finite-domain CSP.
- The solver processes 270,000 puzzles per second, significantly outperforming other methods like Kona 1.0.
- Utilizes intrinsic Riemannian curvature to guide computational resources to regions of highest impact.
- Implements a three-phase CUDA pipeline derived from the Davis Field Equations, automatically routing instances by geometric complexity.
- Classical heuristics (MRV, degree heuristic, checkerboarding) emerge as special cases within the curvature-guided framework.
- The constraint graph forms a genuine discrete Riemannian manifold with measurable curvature invariants.
- A trichotomy parameter Γ classifies instances by geometric complexity, optimizing phase selection without manual intervention.
- Demonstrated on extreme Sudoku puzzles (15-clue, 66 empty cells), solving the hardest instances in under 9ms on consumer GPUs.
- The framework is general-purpose, applicable to any finite-domain CSP expressible with pairwise constraints.
- Named in honor of David H. Blackwell, a pioneer in dynamic programming and game theory, aligning with the solver's decision-making core.