Issue500

Title M&S refactoring: normalize transitions of labels separately
Priority wish Status resolved
Superseder Nosy List malte, silvan
Assigned To silvan Keywords
Optional summary

Created on 2014-12-12.15:41:33 by silvan, last changed by silvan.

Messages
msg3930 (view) Author: silvan Date: 2014-12-13.23:10:52
Merged.
msg3929 (view) Author: silvan Date: 2014-12-13.22:24:07
Yes, I need to clean up. I can simplify the comment, but I'd leave the
mentioning of irrelevant labels in for the moment.

I'll see how large the changes are and then write again.
msg3928 (view) Author: malte Date: 2014-12-13.22:22:25
Great results! :-) I had a quick look at the code; if I looked at it correctly,
there's some leftovers of the old method in normalize that should be cleaned up.

Perhaps we can also simplify the comment at the start of the method to remove
the mention about detecting new irrelevant things so that we don't forget it
later (after merging the issues for that, assuming we'll be able to do that)
when it's no longer relevant?

After the cleanup, I can have another look at the code if you want (for which it
would be best to start a pull request or upload it to Rietveld), but you can
also merge directly if you want.
msg3927 (view) Author: silvan Date: 2014-12-13.22:09:55
v1 implements everything as before, performing the sorting/removal of duplicates
separately for every label.

v2 uses std::sort and std::unique for the same purpose.

v1-baseline results:
http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-12-12-issue500-v1-cggl-comp.html
http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-12-12-issue500-v1-dfp-comp.html
http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-12-12-issue500-v1-rl-comp.html
Results look insanely good.

v2-baseline results:
http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-12-13-issue500-v2-cggl-base-comp.html
http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-12-13-issue500-v2-dfp-base-comp.html
http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-12-13-issue500-v2-rl-base-comp.html
Those are even better!

v2-v1 results:
http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-12-13-issue500-v2-cggl-v1-comp.html
http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-12-13-issue500-v2-dfp-v1-comp.html
http://ai.cs.unibas.ch/_tmp_files/sieverss/2014-12-13-issue500-v2-rl-v1-comp.html
Even the comparison to v1 is nearly always strictly positive.

If you whish to have a look at the code, you find it at my bitbucket repository.
msg3918 (view) Author: silvan Date: 2014-12-12.15:41:33
This is part of meta issue432.

Since the label reduction has been completely split off from normalization,
there is no more need to separate all transitions into buckets simultaneously
when normalizing, constructing them again afterwards, but we can sort and remove
duplicates for every label separately. This hopefully saves memory consumption.
History
Date User Action Args
2014-12-13 23:10:52silvansetstatus: chatting -> resolved
messages: + msg3930
2014-12-13 22:24:07silvansetmessages: + msg3929
2014-12-13 22:22:25maltesetmessages: + msg3928
2014-12-13 22:09:55silvansetmessages: + msg3927
2014-12-12 15:41:33silvancreate