Done.

LGTM.

If you're all happy with it, feel free to merge.

I also integrated the last comments from bitbucket. From my point of view, this
can be merged.

Results look good (pretty much no changes):
http://ai.cs.unibas.ch/_tmp_files/sieverss/issue710-v1-issue710-base-issue710-v1-compare.html

I also addressed most of your comments, in particular writing a hopefully better
documentation for the hill climbing method.

I'm also done with my comments. And yes, I think the 2 ipdb configurations you
mentioned are a good test set.

I left some comments on BitBucket.

I saw that two PDB specialists put themselves onto the nosy list :-) Who is up
for a review? It's not too large:
https://bitbucket.org/SilvanS/fd-dev/pull-requests/25/issue710/diff

I'll also setup some experiments; I assume running "astar(ipdb())" and
"astar(ipdb(max_time=900))" is enough?

When working with the hill climbing generator, I (again) stumbled across some
subtleties and decided to open this issue for a few simplifications. For
instance, I'd like to get rid of keeping both the candidate patterns and the
PDBs around in parallel, rather than just keeping the PDBs from the beginning
on. After all, we can still access the patterns from the PDBs.