Robust statistics and optimmization from Python

[tkpmep]

I use Python to generate a huge amount of data in a .csv file which I
then process using Excel. In particular, I use Excel's solver to solve
a number of non-linear equation, and then regress the results of
hundreds of calls to Solver against a set of known values, enabling me
to calibrate my model. This is a pain: i'd much rather perform all the
computations in Python and improve on Excels' regression as well.

Questions:
1. Is there a way to perform (or make a call to) a non-linear
optimization from Python?
2. Do Python packages for robust statistics (robust regression in
particular) exist. If so, which one would you recommend/

Thanks, as always, in advance for the guidance

Thomas Philips

[Robert Kern]

tk****@hotmail.com wrote:

I use Python to generate a huge amount of data in a .csv file which I
then process using Excel. In particular, I use Excel's solver to solve
a number of non-linear equation, and then regress the results of
hundreds of calls to Solver against a set of known values, enabling me
to calibrate my model. This is a pain: i'd much rather perform all the
computations in Python and improve on Excels' regression as well.

Questions:
1. Is there a way to perform (or make a call to) a non-linear
optimization from Python?
Look at scipy. http://www.scipy.org

In [1]: from scipy import optimize

In [2]: optimize?
....
Optimization Tools
==================

A collection of general-purpose optimization routines.

fmin -- Nelder-Mead Simplex algorithm
(uses only function calls)
fmin_powell -- Powell's (modified) level set method (uses only
function calls)
fmin_cg -- Non-linear (Polak-Rubiere) conjugate gradient
algorithm
(can use function and gradient).

fmin_bfgs -- Quasi-Newton method (can use function and gradient)
fmin_ncg -- Line-search Newton Conjugate Gradient (can use
function, gradient and hessian).
leastsq -- Minimize the sum of squares of M equations in
N unknowns given a starting estimate.

Constrained Optimizers (multivariate)

fmin_l_bfgs_b -- Zhu, Byrd, and Nocedal's L-BFGS-B constrained
optimizer
(if you use this please quote their papers --
see help)

fmin_tnc -- Truncated Newton Code originally written by
Stephen Nash and
adapted to C by Jean-Sebastien Roy.

fmin_cobyla -- Contrained Optimization BY Linear Approximation
2. Do Python packages for robust statistics (robust regression in
particular) exist. If so, which one would you recommend/

Offhand, I can't think of any, but it's easy enough to do maximum
likelihood with Laplacians and the functions above. If you find suitable
FORTRAN or C code that implements a particular "robust" algorithm, it
can probably wrapped for scipy relatively easily.

--
Robert Kern
rk***@ucsd.edu

"In the fields of hell where the grass grows high
Are the graves of dreams allowed to die."
-- Richard Harter

[beliavsky]

Robert Kern wrote:

<snip>

If you find suitable
FORTRAN or C code that implements a particular "robust" algorithm, it
can probably wrapped for scipy relatively easily.

An alternative would be to call R (a free statistical package) from
Python, using something like the R/SPlus - Python Interface at
http://www.omegahat.org/RSPython/ . Many statistical algorithms,
including those for robust statistics, have been implemented in R,
usually by wrapping C or Fortran 77 code.

[araki]

use R. it's pretty highend, and there is an interface for python.

[Tim Churches]

be*******@aol.com wrote:

Robert Kern wrote:

<snip>

If you find suitable
FORTRAN or C code that implements a particular "robust" algorithm, it
can probably wrapped for scipy relatively easily.

An alternative would be to call R (a free statistical package) from
Python, using something like the R/SPlus - Python Interface at
http://www.omegahat.org/RSPython/ .

Unless you really want to call Python from R (as opposed to calling R
from Python), I strongly suggest that you use RPy (http://rpy.sf.net)
rather than RSPython. RPy is much easier to install and use and far less
buggy than RSPython.
Many statistical algorithms,
including those for robust statistics, have been implemented in R,
usually by wrapping C or Fortran 77 code.

Yup, and if you really don't like the extra dependency or extra memory
requirements of R and RPy, it is often possible to port the R code back
to Python (or Numeric Python), with some effort.

Tim C