stephenbrooks.orgForumMuon1Generalan eliminate-runs-that-are-already-done feature?
Username: Password:
Search site:
Subscribe to thread via RSS
[dpc] Immortalsean
2004-10-19 06:32:04
with such an option a computer could check if someone already did the same run.  it could skip that run and do something a little different, so there will be more evolution paths.  this way many computers won't do the same calculations others already did.

maybe somehow it could also be implemented for not-internet-connected computers too.

one way is a large samplefile which should not be performed by the computer, but the computer should perform it just a bit altered and see which gets a better result.  or simply skip it and save some cpu time.  the problem is that the file has to be really big to get more accurate.

another way is to let the computer check an internet server if the run has been performed before, etc. the problem here is that this would cause lots of traffic to these servers.

but it can certainly be an idea.
[DPC]Stephan202
2004-10-19 12:02:32
I think this plan creates a lot of work for Stephen, while it's not really needed.

A) There are many configurations.  The number of doubles seems to be minimal.
There is absolutely nothing wrong with doubles.  Compare it to the situation where a new highest yield is found.  Then a simulation is recalculated at least another four times.
kitsura
2004-10-19 14:34:39
Its almost impossible to come up with duplicate results.  Even running a simulation on the exact same breed you will come up with different results.  So its not necessary.
Nexus
2004-10-29 01:12:58
I think there's a flaw in the quarantine method.  Say the top score so far was 10% and I got a result of 11%. That would be run a further 4 times, but I might only get 8,7,9 and 8 for an average of 8.6%

Yet if my first result was 9.99% it would not be tested again.  But I could well have gotten four additional scores over 10 if the result had been quarantined.
Stephen Brooks
2004-10-29 06:47:06
That's right: the quarantine doesn't eliminate all statistical variation, it just makes sure that the _best_ result at any one time has been run at least 5 times, and this is important because that is the result that will be copied the most.  I could (obviously) make all results be run 5 times but I cosidred the quarantine gives a substantial improvement in the best one without a particularly high price of reculculation.
Nexus
2004-10-29 16:40:34
Obviously you wouldn't test the lowest results again, I just think that if you've got a result that's within a whisker of top place it'd be worth having another look at.
Stephen Brooks
2004-10-30 06:36:44
Yeah, other combinations of rechecking are possible.  If you look at my latest talk in the reports section (November 2004 - I'll give it on Tuesday), you'll see I'm going to mention the 'noise' problem to others at the meeting and see if they have any clever ways of smoothing it out besides just simulating massively more particles.  I've wondered if I should have a 'recheck band' with a certain width within which everything is rechecked; and then there's the other idea of letting the number of rechecks increase until you get a certain desired accuracy.
: contact : - - -
E-mail: sbstrudel characterstephenbrooks.orgTwitter: stephenjbrooksMastodon: strudel charactersjbstrudel charactermstdn.io RSS feed

Site has had 25168308 accesses.