notes

We can probably do single-step global corrections as an extension of
the heuristic class.  We can get parent pointers from the search node,
and compute the average single step error as a result.

Path-based single step corrections require that we associate data with
each node in the search space.  We could do this by expanding the
contents of a search node / search state.  This is probably unfair to
the algorithms that don't require information of that sort.  Instead
we'll have to keep a hashtable inside of the heuristic.

Still not sure how to deal with helpful actions along the frontier.

Need to figure out batch experiment running,

http://www.fast-downward.org/ScriptUsage

talks about how to do this with the internal methods.  It's not going
to produce box-plots or anything other than a set of win counts, but
it's a good start.

Regression shouldn't be too hard considering what we have
implemented.  We just need to take a longer look at what is considered
to be best in terms of the search code.