We discovered by chance that dependencies can be used to track modifications to planners and their environments. We ran Phoenix in one configuration, call it A, and collected execution traces from which we derived a set of significant dependencies, . Then we modified Phoenix slightly-we changed the strategy it used to select failure recovery actions-and ran the modified system and collected execution traces, and, thus, another set of dependencies, . Finally, we added two new failure recovery actions to the original set, ran another experiment, and derived another set of dependencies . To our surprise, the intersections of the sets of dependencies were small. However, both and contained more dependencies than , suggesting that the size of an intersection mimics the magnitude of modifications to a system.
These results are only suggestive, but they raise the possibility that particular planner-environment pairs can be characterized by sets of significant failure-action dependencies. If true, this technique might enable us to identify classes of environments and planners. Today, we assert on purely intuitive grounds that some environments are similar; in the future we might be able to measure similarity in terms of the overlap between sets of dependencies derived from a single planner running in each environment. Conversely, we might measure the similarity of planners in terms of dependencies common to several planners in a single environment.