Advancing Weighted Partial MaxSAT Solving via Differentiable GPU Guidance
Summary
- Researchers introduce a GPU-CPU hybrid framework for Weighted Partial MaxSAT (WPMAX-SAT) that combines CPU solvers with differentiable GPU optimization.
- The GPU stage optimizes instance-specific categorical decision parameters without requiring a training dataset or pretrained predictive model.
- A hierarchical differentiable loss prioritizes hard clause satisfaction before optimizing the weighted soft objective to address clause asymmetry.
- The framework transfers GPU-optimized variable logits to CPU solvers via partial guidance and phase initialization modes.
- Evaluated on MSE 2024 benchmark instances, the hybrid framework improves average solution scores by up to 4.9% over top baseline solvers within a 300-second time limit.
Why it matters
This GPU-accelerated hybrid approach significantly improves Weighted Partial MaxSAT solving efficiency and solution quality, benefiting complex constraint satisfaction and optimization tasks in AI and engineering.