SPLMC / reana-spl

ReAna variants for empirical comparison of analysis strategies.
2 stars 26 forks source link

Assess ADD ordering heuristics #24

Open thiagomael opened 4 years ago

thiagomael commented 4 years ago

I came up with some hypotheses regarding the impact of features in the ordering of ADD variables. For this discussion, keep in mind that we have one (and only one) ADD variable for each feature name in the Feature Model. Also, we are interested in the better ordering for the reliability ADD, not the one representing just the FM rules.

The (non-exhaustive) set of hypotheses is as follows:

  1. Mandatory features seem to be better suited to come first (closer to the ADD root).
  2. Optional features should come last (closer to the leaves).
  3. OR features and Alternative ones should be grouped according to their parenthood in the FM. That is, alternative features that have the same parent should correspond to neighbor variables in the ADD.
  4. Features that occur in the presence condition of an RDG node have an impact in the ADD size that is proportional to the relevance of such node. It is an open question whether this means that such features should come first or last in the ordering.

By "node relevance" we mean how much this node's reliability impacts the overall system. Maybe this could be measured by the node's in-degree, but it is also possible that we need a more advanced metric (such as graph centrality).

The goal of this issue is to validate or refute the hypotheses just described. The valid ones could be used as heuristics to improve analysis time.