Generating diverse solutions to the Boolean Satisfiability Problem (SAT) is a hard computational problem with practical applications for testing and functional verification of software and hardware designs. We explore the way to generate such solutions using Denoising Diffusion coupled with a Graph Neural Network to implement the denoising function. We find that the obtained accuracy is similar to the currently best purely neural method and the produced SAT solutions are highly diverse, even if the system is trained with non-random solutions from a standard solver.
翻译:生成各种解决布利安可满足性问题的方法是一个难以计算的问题,在软件和硬件设计测试和功能核查方面存在着实际应用。我们探索了如何利用Denoising Difmulation和图形神经网络来生成这些解决方案,以实施去除功能。我们发现,获得的准确性与目前最佳的纯神经方法相似,所生成的SAT解决方案也非常多样,即使该系统是从标准求解器得到非随机解决方案的培训。