Mixed-Initiative Workflow Composition

Abstract

Complex applications in many areas, including scientific computations and business-related systems, are represented as computational workflows composed out of multiple components. There are several approaches that help a user compose these workflows. Some composition systems implement a user-system interactive approach (Kim, Spraragen, and Gil 2004). These systems are useful for expressing user preferences during composition; however, they can be tedious to use if a large number of tasks are needed in the workflow, as composition is done one step at a time, manually. Another approach to workflow composition is full automation, which aims to eliminate unnecessary user interaction during composition (Blythe et al. 2003), (Myers et al. 2003). This approach is efficient, but is not ideal if user preferences need to be expressed during composition. Our approach combines the strengths of manual and automatic approaches into mixed-initiative workflow composition. This combined approach uses automated planning techniques, while also incorporating user preferences during composition. Our approach is implemented in a new system, AutoCAT, by combining an interactive workflow editor (Composition Analysis Tool or CAT) (Kim, Spraragen, and Gil 2004) and a planner, Prodigy (Veloso et al. 1995). CAT enables users to manually compose workflows by adding and removing tasks, and by linking the tasks together in order to supply data (i.e., linking task outputs) to the workflow’s user-defined end results. Using AI planning principles, CAT checks the workflow for possible composition errors introduced by a user. For instance, the workflow’s links are checked for correctness (e.g., “flight number” task output linked to “driver license number” task input is incorrect), and checks for unnecessary workflow tasks (for example, if the workflow’s only end result is “flight reservation”, then a “real estate service” task is unnecessary). CAT’s approach also involves searching a knowledge base of task types in order to suggest to the user the next composition steps that should be taken (for _______________________