[SciPy-dev] Nonlinear constrained optimizer in SciPy

Travis Oliphant oliphant at ee.byu.edu
Fri Apr 9 20:49:46 CDT 2004


eric jones wrote:

> David M. Cooke wrote:

Thanks to David Cooke and the very informative Jean-Sebastien Roy SciPy 
now has our first multivariable constrained optimization function

Just in time for the Scipy 0.3  release. 

The Zhu, Byrd, Nocedal  L-BFGS-B  optimizer is named

scipy.optimize.fmin_l_bfgs_b

It uses the same interface David constructed.  I removed dependency on 
linpack and blas and use instead the lapack returned from scipy_distutils.

The docstring contains the information for the papers people should 
quote if they use the algorithm in their work.

I notice that it is slower for unconstrained problems than the other 
algorithms which I suppose is not surprising.  What have other people 
experienced?

I just did a simple test on the rosenbrock function (notice that 
optimize.rosen and optimize.rosen_der provide the function and gradient 
for this famous function)  both of these are vector-length agnostic so 
that the initial guess defines the order of the problem.  I tested order 
3-5 on windows and linux without problems.

Thanks again,

-Travis Oliphant








More information about the Scipy-dev mailing list