Faster Constraint Solving Using Learning Based Abstractions

This work addresses the problem of scalable constraint solving. Our technique combines traditional constraint-solving approaches with machine learning techniques to propose abstractions that simplify the problem. First, we use a collection of heuristics to learn sets of constraints that may be well abstracted as a single, simpler constraint. Next, we use an asymmetric machine learning procedure to abstract the set of clauses, using satisfying and falsifying instances as training data. Next, we solve a reduced constraint problem to check that the learned formula is indeed a consequent (or antecedent) of the formula we sought to abstract, and finally we use the learned formula to check the original property. Our experiments show that our technique allows improved handling of constraint solving instances that are slow to complete on a conventional solver. Our technique is complementary to existing constraint solving approaches, in the sense that it can be used to improve the scalability of any existing tool.