FoMo - Formula and Model Generation for Learning-Based Formal Methods

Colin Shea-Blymyer, Houssam Abbas

Keywords: Neural networks, Verification, model checking, temporal logics, datasets, formal methods, neurosymbolic reasoning

TL;DR: Generate data to train a neural network for jointly embedding formulas and system models.

Abstract: This paper presents a tool that can generate system model and temporal logic formula pairs such that the formula is satisfied or unsatisfied as desired. Large amounts of independent and identically distributed data is needed for the development of machine learning approaches in formal methods. However, existing data sets of system models and specifications were hand-designed to challenge existing algorithms for formal methods, and so these data sets represent only a small subset of problems. Learning-based approaches to formal methods require training data drawn from a broader distribution. In this work we introduce a tool called ``FoMo'' (for ``Formula and Model'') that generates system models from graph distributions and can sample a specification language to generate properties. FoMo includes functions to generate pairs of formulas and satisfying or unsatisfying system models, and to generate traces from systems. The tool has features for working with linear temporal logic, and with weighted automata. We demonstrate the use of this tool by training a neural network to jointly embed formulas and models to allow for model checking as a classification task. The capabilities offered by FoMo allow researchers to generate system models and properties according to many research needs, and helps provide machine learning systems with generated, multi-modal data sets.