stephenbrooks.orgForumMuon1Generaltop100 suggestion
Username: Password:
Search site:
Subscribe to thread via RSS
AlArenal
2002-06-04 12:35:11
Hi Stephen!

Now you have integrated automatic upload of results, wouldn't it make sense to automatically download the top100?  I guess you already have a script to extract the top results out of all uploaded results or could easily write yourself such a script.
Stephen Brooks
2002-06-04 15:32:54
If I upload the top100 before most people are prodcuing results with 4.1x, we might just end up with everyone's program optimising in the same place, which would be a waste.

I like them working independantly for a while and then combining designs at a much later stage.


"As every 11-year-old kid knows, if you concentrate enough Van-der-Graff generators and expensive special effects in one place, you create a spiral space-time whirly thing, AND an interesting plotline"
gogomaus
2002-06-05 05:31:49
Stephen,
agree to not offering the top100 until the _final_ stage (never before).  On the other hand a lot of newcomers are wasting their time and PC capacity on evolving mediocre results.
I made a proposal some days ago in another thread, but got no response, so here´s the second chance :
Offer seed packages of some 50 good results (>1.6% or so), randomly extracted out of all !  By this way the community gen-pool could be mixed/enriched smoothly without loosing diversity.  (has successfully been used under other distributed calculations with genetic approach).
kruemi [SwissTeam.NET]
2002-06-06 04:51:56
Hm... Another way could be, even if a lot of old results exist, just do some fully-random calculations between the others!  Lets say, evvery 5th try is fully random, without the genetic algorythm...

bye

kruemi
gogomaus
2002-06-06 07:37:08
quote:
Originally posted by kruemi [SwissTeam.NET]:
Hm... Another way could be, even if a lot of old results exist, just do some fully-random calculations between the others!  Lets say, evvery 5th try is fully random, without the genetic algorythm...
bye kruemi


This is already done !!!  25% of all runs are using randomly generated parameters.  But it is still pure lottery for every user to catch a really good result.  We face the situation of user´s evolution on an island with no connection to genetic efforts on other islands.  This is of course a valid approach of evolution, but definitely not the fastest and most efficient one.
Stephen Brooks
2002-06-07 10:45:58
Aha you've given me an idea.  There's now a very large results.dat file on the web page containing everyones v4.1 results so far.  When this gets much bigger (e.g. 4M I might also have a version up there that's condensed a bit (some rubbish results removed).


"As every 11-year-old kid knows, if you concentrate enough Van-der-Graff generators and expensive special effects in one place, you create a spiral space-time whirly thing, AND an interesting plotline"
gogomaus
2002-06-10 00:47:19
Oh no,
now you did it !  Offering all results will lead exactly to what you did not want (and you denied it 3 days ago !).  Everyone using this results.dat file (and the "buggy" v4.11 with "single" parents) will create repeated calculations of the best runs done before.
I gave tips enough, but you prefer to ring in the final round just now.

Please give a clear hint on the website and within the forum that you changed the evolution of muon program, requesting people to download and use that new version.  Maybe at least some diversity can be achieved in future (if you like to do so).
kruemi [SwissTeam.NET]
2002-06-10 04:09:35
Dverity is given by the 25% completely-random Calculations and in the Fact that not everyone is downloading the big results.dat-file!
I thin it's a good idea to try this way!
I have two clients running on my machine... one, which i started without results.dat and one which i started with some results...
The differences in the results are huge!

cu

kruemi
gogomaus
2002-06-10 04:34:05
@ kruemi
random calculations do not really contribute to diversity of _good_ results, because there is only a very little chance to find a result >> 2% by chance !!!
All other results will be considered as garbage and never be chosen again for genetic crossover / linear combinations, as long as the program sticks to only _one_ parent (and this will be only selected out of the very best of all before !).

I know, you were a very lucky user by finding a good result in the early stage, but all the others had to wait for a long time and will be keen on super seeding.

Now most evolution work will focus on extremely small, distinct cut out of what would be worth to investigate.  I believe this is not a good idea.... wink
Stephen Brooks
2002-06-13 15:24:41
When the data set is very large (like in this one), the distribution of how the parents are chosen is actually more spread out.  When you've got 100 results it'll stick to only a dozen or so of the better ones, but with this many (thousands) it will choose parents from all sorts of different ones.


"As every 11-year-old kid knows, if you concentrate enough Van-der-Graff generators and expensive special effects in one place, you create a spiral space-time whirly thing, AND an interesting plotline"
gogomaus
2002-06-14 02:31:26
Still not convincing.
As the crossover/linear combination gen tool (too) often selects identical parents, creation of clones is considerably high.  So, in case of 15000 base results there may be up to 2000 "potential parents ", but in fact only some dozens or hundreds different one !  [This bug should be fixed.]
That is nothing compared to the billions of good solutions on the 14-D-surface...
We would investigate only (around) some spots in a huge space, if not the mutation part would offer some alternatives.
Personally, I started with nothing and after 150 runs I had 3 different result types with about 1.5%. Then I seeded 20 carefully selected, divertified results from SwissTeam (1.7-1.9%).  After 200 additional runs my results.dat was flooded again with clones/very similar parameters of only 5 parents.  Looks like the other 15 never have been touched !?
Could you explain, how the selection criteria for "parents" (should) work ?
Does it have a lower result limit ?  Does it start from top down ?
Would be nice to have more diversity... roll eyes
: contact : - - -
E-mail: sbstrudel characterstephenbrooks.orgTwitter: stephenjbrooksMastodon: strudel charactersjbstrudel charactermstdn.io RSS feed

Site has had 25160858 accesses.