Abstract.Many studies have shown that selective undo, a variant of the widelyimplemented linear undo, has many advantages over the prevailing model. In this paper, we define a task model for implementing selective undo in the face of dependencies that may exist between the undone action and other subsequent user actions. Our model accounts for these dependencies by identifying other actions besides the undone one that should also be undone to keep the application in a stable state. Our approach, which we call cascading selective undo, is built upon a process-programming language originally designed in the software engineering community. The result is a formal analytical framework by which the semantics of selective undo can be represented separately from the application itself. We present our task model, the selective undo algorithm, and discuss extensions that account for differing kinds of inter-action dependencies.

While various models of undo have been proposed over the years, no empirical study has yet been done to discover which model of undo most closely aligns with what users expect an undo command should do. In this paper, we discuss the results of such a study that compares the ubiquitous linear undo model with two variations of selective undo: script selective and cascading selective. Unlike the script model, cascading selective undo takes into account dependencies between user actions. Our study shows that, for the application studied, when a user is asked to perform undo in the absence of any guidance, the user will tend to gravitate toward an undo mechanism that uses existing dependencies between user actions. Specifically, we show that subjects prefer the dependency-aware aspects of cascading undo over either linear or script selective undo.