Automated Backend Selection for ProB Using Deep Learning

Analysing ProB’s Constraint Solving Backends

“…In previous work [19] we used convolutional neural networks, but we were unable to extract the knowledge accumulated by the classifciation due to the black box nature of the neural networks. Hence, we started to use decision trees, as one can easily extract classification rules after the training phase.…”

“…encountered during symbolic verification. In previous work [18,19], we trained neural networks to decide for a given constraint which backend should be used. We compared two approaches: one based on feature vectors derived from domain knowledge, and one based on encoding constraints as images.…”

Towards a Shared Specification Repository

“…Other use cases, such as data validation [7] work by executing a model along one particular, linear path, while others, like constraint solving problems, sometimes work on machines without variables, consisting of a single state. Most recently, machine learning (ML) techniques are applied to model checking or synthesis as well, and require a large number of specifications, e.g., in order to extract and re-combine predicates [6]. Even with access to numerous machines, it is time-consuming and cumbersome to identify machines to use for benchmarking, especially since only a small amount of data can be presented in a typical research article.…”

“…https://sv-comp.sosy-lab.org/2020/ 5. Which inspired the second author to generate another library, Dozens of Problems for Partial Deduction https://github.com/leuschel/DPPD 6. Cf.…”