Workflow Discovery with Semantic Constraints: The SAT-Based Implementation of APE

Abstract

Science today is increasingly computational, and many researchers regularly face the need of creating purpose-specific computational pipelines for their specific data analysis problems. The manual composition and implementation of such workflows regularly costs valuable research time. Hence, many scientists wish for a system that would only require an abstract description of their intended data analysis process, and from there automatically compose and implement suitable workflows. In this paper we describe APE (the Automated Pipeline Explorer), a new implementation of a synthesis-based workflow discovery framework that aims to accomplish such automated composition. The framework captures the required technical domain knowledge in the form of tool and type taxonomies and functional tool annotations. Based on this semantic domain model, the framework allows users to specify their intents about workflows at an abstract, conceptual level in the form of natural-language templates. Internally, APE maps them to a temporal logic and translates them into a propositional logic instance of the problem that can be solved by an off-the-shelf SAT solver. From the solutions provided by the solver, APE then constructs executable workflow implementations. First applications of APE on realistic scientific workflow scenarios have shown that it is able to efficiently synthesize meaningful workflows. We use an example from the geospatial application domain as a running example in this paper.