Efficiently generating structurally complex inputs with thousands of objects

We present Shekoosh, a novel framework for constraint-based generation of structurally complex inputs of large sizes. Given a Java predicate that represents the desired structural integrity constraints, Shekoosh systematically explores the input space of the predicate and generates inputs that satisfy the given constraints. While the problem of generating an input that satisfies all the given constraints is hard, generating a structure at random, which may not satisfy the constraints but has a desired number of objects is straightforward. Indeed, a structure generated at random is highly unlikely to satisfy any of the desired constraints. However, it can be repaired to transform it so that it satisfies all the desired constraints. Experiments show that Shekoosh can efficiently generate structures that are up to 100 times larger than those possible with previous algorithms, including those that are based on a dedicated search and also those that use off-the-shelf enumerating SAT solvers.