RandCSPBench
收藏资源简介:
RandCSPBench是由博科尼大学等机构联合构建的约束满足问题基准数据集,包含3-SAT、4-SAT、3-col和5-col四类NP难问题实例,总计41.2万条数据。数据集通过调节子句变量比(α)和约束密度(c)等参数,系统生成不同难度级别的实例,覆盖从可满足性阈值附近到高硬度区间的完整谱系。其创新性在于引入统计物理相变理论指导的渐进式难度设计,并首次纳入K>3的高阶问题。该数据集旨在评估图神经网络与传统算法在组合优化问题中的性能边界,为算法鲁棒性研究提供标准化测试平台。
RandCSPBench is a benchmark dataset for constraint satisfaction problems jointly constructed by Bocconi University and other institutions. It includes four categories of NP-hard problem instances: 3-SAT, 4-SAT, 3-col, and 5-col, with a total of 412,000 data entries. The dataset systematically generates instances of varying difficulty levels by adjusting parameters such as the clause-to-variable ratio (α) and constraint density (c), covering the full spectrum from regions near the satisfiability threshold to the high-hardness regime. Its key innovation lies in adopting a progressive difficulty design guided by the statistical physics phase transition theory, and it is the first benchmark dataset to incorporate high-order constraint satisfaction problems with K>3. This dataset aims to evaluate the performance boundaries of graph neural networks and traditional algorithms on combinatorial optimization problems, providing a standardized testbed for research on algorithm robustness.
- 1Benchmarking Graph Neural Networks in Solving Hard Constraint Satisfaction Problems博科尼大学·数据科学与分析研究所; 佛罗伦萨大学·物理与天文系; 罗马第一大学·物理系; 意大利国家研究委员会·纳米技术研究所; 哈瓦那大学·理论物理系 · 2026年



