A small addendum to the comment in msg4101: during the work on issue521, I found
that the constructor of composite transition systems is a place where changes in
the order in which labels are stored in a label group can occur, because we use
an unordered_map to split label groups of one component along the groups of the
other. This then may again influence the resulting bisimulation when computed on
the composite.

Done.

If you like it, please merge.

After adding another small fix, all other configurations except the one we
already saw in the previous comparisons behave exactly the same as before:

http://ai.cs.unibas.ch/_tmp_files/sieverss/2015-03-19-issue516-v2-v4-comp.html

To precise my previous statement: only the order in which bisimulation considers
the label groups matters. If, for instance, we always sorted the label groups,
also the results of the baseline would change, and no more differences between
the baseline and the latest version would exist.

I think this can be merged.

Works for me.

V3 compared against the fixed baseline V2 for DFP can be found here:

http://ai.cs.unibas.ch/_tmp_files/sieverss/2015-03-17-issue516-v3-comp.html

The small changes in parcprinter are due to a change in the order in which
labels are considered during label reduction, changing the order in which
transitions of new labels occur in transition systems being shrunk with
bisimulation, which apparently depends on the order of label-transition groups.
As the changes are so small, I suggest to live with it. I will next run a few
other configurations before calling this one resolved.

Another issue came up during the implementation of this issue, which is a bug
which caused dead labels to possibly not be grouped together when computing a
composite transition system. Fixing this slightly improved construction time
(and memory consumption) of the merge-and-shrink process.

http://ai.cs.unibas.ch/_tmp_files/sieverss/2015-03-17-issue516-v2-comp.html

This is part of meta issue432.

We would like to call store transitions of labels that are "locally equivalent"
with the exception of their cost. We can then associate the minimum cost of all
labels of a group for purposes such as label reduction and bisimulation.