|
|
I am working on a project to review possible solver technologies, we currently use a Levenberg Markov - Least squares implementation - to minimise a non-linear objective function with 6 dimensions.
I am new to optimization, but have been playing about with the CompactQuasiNewton Solver in MSF which seems to be the most appropriate for Non-linear problems. I have also read that the interior point solver within SFS may handle such problems.
Any advice/experiences of similar applications would be most appreciated.
Cheers Paul
|
|
|
|
Hi Paul,
In general the CQN solver is the best path for Nonlinear Least Squares Optimization.
For documentation see Documents\MSF-SolverProgrammingPrimer.docx in your installation directory. There is a C# sample in “User Documents directory\Microsoft Solver Foundation\Samples\Solvers\C#\NLP”
Our Interior Point solver can handle Linear Least Squares, as the function to optimize is quadratic. In this case you can program against the Solver Foundation Services API, which is easier than coding against solver directly.
-Shahar
|
|
|
|
Hi Shahar, Thanks for the response. Do you know if there are any examples of using the CompactQuasiNewtonSolver with multiple variables defined in x? I have defined an additional variable, set the number of variables to 2 and altered the sinex_value and gradient functions to optimise a function with two variables x 0 and x 1, but it only adjusts the x 0 variable on each iteration. Leaving x 1 at the seed value. I'm probably doing something wrong, any pointers most appreciated. Cheers Paul
|
|
|
|
Hi Paul,
Can you please share the code?
Are you sure x1 starting point is not a local optimum?
Thanks,
Shahar
|
|
|
|
Hi Shahar,
Have got past my initial issue with the gradient vector, my partial derivatives were incorrect. Will keep you posted as things develop.
Cheers Paul
|
|
|
|
|