Topic: Simulated annealing (scipy.optimize.anneal)
Am I correct to say that it wouldn't be too difficult to connect scipy's anneal to OO? What is the best way to get involved in doing this?
Re: Simulated annealing (scipy.optimize.anneal)
Hi Danie,
I had informed scipy developers for twice if not for thrice, that their anneal() implementation has serious drawbacks.
See for example this post.
Today I had check it once again and I have obtained the same results:
from numpy import asfarray
from scipy.optimize import anneal
x_opt, retval = anneal(lambda x: (x**2).sum(), asfarray([1,-1]))
print x_opt
# and typical output is Tooo far from exact solution (0, 0):
array([ 4.63939483, -12.77243753])
(it is not constant since random numbers are used).
So taking all above into account, I have no willing to have it connected into OO, at least until scipy dev team will have it fixed.
OpenOpt already has several excellent global solvers.
Regards, D.
Re: Simulated annealing (scipy.optimize.anneal)
Hi Dmitrey
Unfortunately the link appears to be broken. I am not all too familiar with the inner workings of the anneal implementation. But I believe the default parameters (dwell, T0) are not adequate. At least, I think that these parameters must be chosen wisely depending on the problem at hand. I get this solution (or equivalent) consistently with T0=0.4 and dwell=5000:
[ -2.63230025e-08 1.37166546e-08]
They are certainly not [0 0], but acceptable. My parameter choice was a quick guess and I'm sure that better/faster parameters exist.
I would therefore guess that the drawbacks of the scipy anneal() is in a way intrinsic to the algorithm and not necessarily to the implementation. I have been able to use the scipy implementation with success, but I must admit that guessing the parameters is somewhat of a process.
I could be seriously in error.
Regards, D
Re: Simulated annealing (scipy.optimize.anneal)
It's not a broken link, it's server scipy.org that is down already for several hours, and this is one of the reasons why OpenOpt had migrated to another host - the scipy.org server crashes too often.
I don't want to discuss scipy programmers approach, but as for me I DO try to care for quality - both OO kernel and solvers connected.
Of course hand-turning parameters can somehow improve solver results, but a good enough solver should yield a good results for all-default settings.
As for the parameters T0=0.4 and dwell=5000, for the puny problem sum(x^2) with 2 variables it takes 5 sec for x_opt = array([ 0.19294977, -0.21508007]) and ~25 sec for x_opt = array([ 1.34310712e-08, -1.06388572e-08]) (I have AMD 3800+ X2). The times are some thousands times greater in comparison with other global solvers connected to OO, and with essential nVars >> 1 it will differ even more.
So connecting anneal is inappropriate for now, it will be just a waste of time both for me and for OO users.
Re: Simulated annealing (scipy.optimize.anneal)
Makes sense.
Thanks
D
Re: Simulated annealing (scipy.optimize.anneal)
Hi,
Does anyone know where I can find the source code for a simulated annealing algorithm? The basic procedures (generate and search) for temperature and cooling ratio would be nice, but of course the code for a labour scheduling/personel rostering problem in a manufacturing environment would be great because this is the problem I'm trying to solve.
I am doing everything in C++ by the way 
Thanks!
Re: Simulated annealing (scipy.optimize.anneal)
Today 11:01:52 wrote:
Does anyone know where I can find the source code for a simulated annealing algorithm
Today 11:01:52 wrote:
I am doing everything in C++ by the way
google "simulated annealing c++" yields 36000+ entries, I guess you'll get it there.