More Complex Dependencies.

The dependencies examined so far have been limited to temporally adjacent failures, actions or the combination of each. The combinatorial nature of dependency detection precludes arbitrarily long sequences of precursors. More complex dependencies can be discovered either by controlling the collection of data to selectively test for particular dependencies (through experiment design) or by heuristically controlling the construction and comparison of dependencies (through enhancements to the Heterogeneity G-Test). A new experiment design would selectively eliminate actions from the available set to test whether each precipitates or avoids particular failures (i.e., an ablation or lesion study). Rather than examining all possible chains of which some action is a member, the new analysis removes the action from consideration, which results in execution traces free from the interaction of the missing action. Dependency sets from the different execution traces would be compared to assess the influence of the missing action.

Alternatively, dependency sets can be built iteratively from subsets; the Heterogeneity G-Test suggests a method of doing so for singletons and pairs, but cannot be applied in a straightforward fashion to longer combinations. We need to enhance the technique to compare longer combinations and use the results of comparing sets of shorter precursors to motivate the search for longer ones. For example, if some singleton subsumes a set of pairs, it seems unlikely to be necessary to look at longer combinations beyond the pairs. In effect, Heterogeneity testing becomes a means of controlling heuristic search through the potentially combinatorial space of possible dependencies.