This repository contains algorithms I'm preparing for scipy as part of my GSoC 2015 work. In order to use: clone this repository and add the path to the folder to PYTHONPATH. Tested in Python 3.4 and 2.7 with numpy 1.9.2 and scipy 0.15.1.

The algorithms is contained in package bounded_lsq.

1. leastsqbound.py is a wrapper over scipy.otpimize.leastsq which does bounded-to-unbounded variables transformation. It is the exact copy from here, master branch, taken on 27 June 2015, commit 937c67500cf9340c31d5cfbbb274894e3b59bd89. This is the fixed version and now it works well.
2. dogbox.py implements a dogleg trust-region algorithm applied to a rectangular trust region. You can read the description in my blog.
3. trf.py implements a special Trust Region Reflective algorithm which combines several ideas. See the description in my blog.

Run benchmarks from benchmarks/run_benchmarks.py. The default usage

python run_benchmarks.py

runs all benchmarks with default tolerance parameters, analytical Jacobian and prints to stdout.

The following example command run only bounded problems with custom tolerance settings using 2-point numerical Jacobian approximation and prints the result into a file:

python run_benchmarks.py report.txt -b -ftol 1e-12 -xtol 1e-12 -gtol 1e-8 -jac 2-point 

Run python run_benchmarks.py --help to see full parameters signature.

For more information about this benchmarks read my post. Your results can be somewhat different to ones reported in the blog, because I keep adjusting algorithms.