Research

Composing computational workflows is essential in many areas, including scientific computations and business-related web services. A new kind of science is emerging from the integration of models developed by individual scientists and groups: end-to-end scientific applications that result from the composition of those individual models. Workflow composition poses challenges for an unassisted user, such as keeping detailed track of how large numbers of complex tasks can be joined together. At the other extreme, though a purely automatic planner can do composition, there may be a large number of equally valid workflows to compute and analyze. Pruning the search might be enhanced by stepwise interaction with the user, who can specify preferences as needed.

Our approach to interactive workflow composition incorporates 1) knowledge-rich descriptions of the individual components and their constraints; 2) a formalism based on well-defined semantics from known planning paradigms; and 3) a formal algorithmic understanding of partial workflows, based on AI planning techniques. Using this approach, we have developed and implemented the Composition Analysis Tool (CAT) system, an ontology-backed, mixed-initiative approach to solving the task composition problem. CAT allows the user to add and link domain tasks at will, but also includes a systematic stepwise algorithm for error checking and fix suggestion generation. CAT can analyze a partial workflow composed by the user, notify the user of issues to be resolved in the current workflow, and suggest to the user what actions could be taken next. If taken, CAT's suggestions help the user compose a workflow that the system considers correct, with the user's preferences being expressed throughout the process.

Screenshot of implemented CAT system: