https://bugs.freedesktop.org/show_bug.cgi?id=80647
Priority: medium
Bug ID: 80647
Assignee: [email protected]
Summary: request to add Nelder Mead simplex to SOLVER for small
dimension non-linear problems
Severity: enhancement
Classification: Unclassified
OS: Windows (All)
Reporter: [email protected]
Hardware: x86 (IA32)
Status: UNCONFIRMED
Version: 4.2.4.2 release
Component: Spreadsheet
Product: LibreOffice
I have been trying to use the solver in Calc to find the minimum of a
non-linear function calculated in the spreadsheet found here:
http://claub.net/temp/Solver-crashes-CALC.xls
The optimized quantity is a sum-of-squares error, cell A134, and the value in
this cell should fall to 10 or less. I set up the solver to find a minimum by
manipulating the values in cells F13 through F17. The initial values I choose
are F13=1.0 and the other set to zero. I am using Calc under Windows 8, version
4.2.4.2, Build ID: 63150712c6d317d27ce2db16eb94c2f3d7b699f8.
I have tried both the DEPS and SCO algorithms. The solver runs for what seems
like a long time (more than a minute on my machine) and makes some progress
reducing the error. Then it, and calc, crash and close. This happens every
time.
What I find strange is that I can run this exact same problem with the same
initial conditions using the Excel solver (GRG nonlienar algorithm) and it
finishes in about 1-2 seconds and never crashes.
The Calc SOLVER really should include a basic, Nelder-Mead simplex based,
non-linear optimization option. This works very reliably for non-pathological
optimization problems of modest dimension (e.g. less than 15-20 variables). I
have applied that algorithm to the problem in the spreadsheet linked above with
100% success for instance, and it would be much faster than these bizarre
evolutionary algorithms for my 5-variable problem. I have some experience
coding the Nelder-Mead algorithm and know of some slight modifications from the
literature that help it to work better than the original implementation and
that keep it from crashing in the case that the simplex collapses or a
duplicate point (within machine precision) is created during the minimization.
Please contact me for details here:
http://audio.claub.net/contact_form.html
Adding the (very reliable) Nelder Mead simplex algorithm to the available
non-linear optimization engines in the Solver would round out its capabilities
nicely and should not be difficult to code (the algorithm is not complicated).
Thanks,
Charlie
--
You are receiving this mail because:
You are the assignee for the bug.
_______________________________________________
Libreoffice-bugs mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/libreoffice-bugs
