Merged and pushed.

Sounds good to merge then. I guess it would make sense to do something about
running out of memory at some point, but that shouldn't block this issue.

Here are the experimental results for iPDB with a 900s time limit:

http://ai.cs.unibas.ch/_tmp_files/seipp/issue768-v1-opt-issue768-base-issue768-v1-compare.html
http://ai.cs.unibas.ch/_tmp_files/seipp/issue768-v1-opt-total_time-ipdb-issue768-base-issue768-v1.png
http://ai.cs.unibas.ch/_tmp_files/seipp/issue768-v1-opt-memory-ipdb-issue768-base-issue768-v1.png

Hill climbing often runs much faster with this patch. The difference in memory usage for some tasks 
probably stems from different pattern collections being generated.  

Coverage decreases by 3 in airport due to running out of memory during hill climbing. This is due to 
the ~2x increase in speed during hill climbing which allows generating more patterns. Coverage 
increases by 1 in logistics00 and parcprinter-08-strips.

The code change is local enough that I'm happy to defer to Florian's review, so
no review needed from my side.

Can we see experimental results?

The code looks all right to me.

Sorry about that. You should have access now.

I got an "Access Denied" message.

I made a pull request at https://bitbucket.org/jendrikseipp/downward/pull-requests/80 .

This change should speed up the hill climbing algorithm.