Feature-Free Approach for SAT Solver Selection

ICLR 2026 Conference Submission17229 Authors

Keywords: Algorithm Selection, Satisfiability Problem, Machine Learning, Graph Embedding, Deep Forest

Abstract: Boolean Satisfiability Problem is a cornerstone in computer science and artificial intelligence, underpinning numerous applications through its ability to solve complex computational problems. However, existing SAT solvers face significant limitations, including the complexity and domain expertise required for feature design, the static nature of many feature sets that limit adaptability to evolving problem structures, and the poor generalization of handcrafted features to new instances, thereby constraining performance across diverse SAT problem distributions. To address these challenges, we introduce a Feature-Free SAT Solver Selection (F2S3), which integrates the Sensitive-Associative Cascade Forest (SACF), Correlation Refinement Factor Graph (CRFG), and Dual-Proximity Graph Representation (DPGR) to address the complexities of SAT problems. F2S3 method transforms problem instances into graph data, utilizes CRFG to maintain the higher-order nature of the graph structure and node relationships, and uses DPGR to enhance the graph data features and map them to low-dimensional vectors. This approach effectively captures the structural intricacies of graph data and improves feature representation in low-dimensional spaces, overcoming the limitations of previous methods regarding feature sparsity and generalization ability. Experiments conducted on datasets from the ASlib database demonstrate that F2S3 outperforms existing solutions, particularly in scenarios where previous methods were hindered by challenges such as feature sparsity and computational inefficiency. The method's performance is evaluated across multiple competitive datasets, showing high gap values and consistent robustness.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 17229