Towards Effective Deep Learning for Constraint Satisfaction Problems

Many attempts have been made to apply machine learning techniques to constraint satisfaction problems (CSPs). However, none of them have made use of the recent advances in deep learning. In this paper, we apply deep learning to predict the satisfiabilities of CSPs. To the best of our knowledge, this is the first effective application of deep learning to CSPs that yields \(>99.99\%\) prediction accuracy on random Boolean binary CSPs whose constraint tightnesses or constraint densities do not determine their satisfiabilities. We use a deep convolutional neural network on a matrix representation of CSPs. Since it is NP-hard to solve CSPs, labeled data required for training are in general costly to produce and are thus scarce. We address this issue using the asymptotic behavior of generalized Model A, a new random CSP generation model, along with domain adaptation and data augmentation techniques for CSPs. We demonstrate the effectiveness of our deep learning techniques using experiments on random Boolean binary CSPs. While these CSPs are known to be in P, we use them for a proof of concept.