[gecode-users] Analysing search performance

[Christian Schulte]

Always use recomputation: the default settings for Gecode's search engine
sets the copying distance (c_d) to 8 and the adaptive distance (a_d) to 2!
Recomputation is done in a rather smart way, so it will not hurt big
perfomancewise.

As itc omes ICL_VAL for distinct, it does the obvious: wait until a variable
ceomes assigned and then remove that value from all other variables. In
problems where one has little propagation anyway, this can work well.

Christian

--
Christian Schulte, http://www.imit.kth.se/~schulte/ 

-----Original Message-----
From: users-bounces at gecode.org [mailto:users-bounces at gecode.org] On Behalf
Of Lars Otten
Sent: Friday, March 03, 2006 1:57 PM
To: users at gecode.org
Subject: Re: [gecode-users] Analysing search performance


On 03/03/06 01:51, Lars Otten wrote:
> The biggest instance I'm currently testing is eating about 1 GB of 
> memory when run, so on my machine it is an issue, yes. ;) I am not 
> using recomputation at the moment (for direct comparison with my 
> randomized search algorithm, which don't use recomputation either), 
> but as I found that mostly increases memory consumption anyway...

Hm, actually it does not seem to be quite that simple. Memory consumption
for dfs-reco always seems to be lower, but wrt. running time things
sometimes get better, sometimes worse (compared to dfs-copy).

Is there any rule of thumb when to apply dfs-reco instead of dfs-copy?

> Just tonight, however, I have made some progress by toying with the 
> ICL for the distinct propagators that are posted. Especially in 
> cooperation with some of the random stuff I implemented, a solution 
> seems to be found rather quickly for most instances -- I'm quite happy 
> about that at the moment! :))

Ok, so I think I understood the conceptual difference between ICL_DOM and
ICL_BND -- in consequence ICL_BND is 'less accurate' but potentially
propagates faster. But where does ICL_VAL fit into that? I am aware that
these levels are only meant to be guidelines for the propagators, still it
would be interesting to know...

Regards,
/Lars

_______________________________________________
Gecode users mailing list
users at gecode.org https://www.gecode.org/mailman/listinfo/gecode-users