Wrapping up

So, the GSoC period has come to an end now.

Project Goals  

My project was about creating Matlab compatible wrappers for the optim package.  Here is a brief list of my project goals.

1- lsqnonlin wrapping nonlin_residmin and residmin_stat
2- lsqcurvefit wrapping nonlin_curvefit and curvefit_stat
3- nlinfit wrapping nonlin_curvefit (it was initially decided to wrap leasqr but changed to avoid extra computations)
4- quadprog wrapping __qp__ instead of qp and returning lambda in the form of a structure as in Matlab
5- fmincon wrapping nonlin_min
6- Test and demos for the above functions
7- Stretch goals: previously decided to create other missing functions or perhaps additional backends but before midterm I decided instead to include optimoptions in my to do list.

Achievements

The functions lsqnonlin, lsqcurvefit and nlinfit are complete with tests and demos and integrated in the optim package. Since nlinfit is from the statistics package in Matlab, additional functions such as statset, statget were required for handling options. These functions are implemented with minor modifications in optimset, optimget and __all_opts__ as statset, statget and __all_stat_opts__ and are now a part of optim package.

The function quadprog required directly wrapping __qp__ instead of qp for the ordering of lambda. It is in the final stages of review and will soon be integrated.

fmincon has not been thoroughly reviewed yet. I will send it to Olaf after quadprog is committed to optim. 

Hiccups in the stretch goal

I couldn't create optimoptions in the GSoC time frame because it was a bit open ended and I had to come up with an object oriented design for the function. I was trying to understand how Matlab implements it for quite some time. Anyway, I didn't pursue it further and shifted my focus on the refinement of my almost complete functions to get them integrated in optim.  

Interesting Takeaways

This is my first experience of working with any open source organization and it's definitely a pleasant one. It's delightful to see people using my functions and possibly benefiting from my work [2-3]. :)

I think I have managed to meet all the goals I set before the start of GSoC. (Regrets? Well, I could have saved more time for optimoptions and it would've been better to discuss it way before than being stuck for a while.)

I'm extremely grateful to the Octave community, especially my mentors Olaf Till and Nir Krakauer for their unrelenting support. GSoC wouldn't have been possible without their constructive feedback. I have learned a lot from this experience.

[1] https://github.com/AsmaAfzal/octave_workspace/issues/1

[2] http://octave.1599824.n4.nabble.com/GSoC-2015-Optimization-Package-Non-linear-and-constrained-least-squares-lsqcurvefit-lsqlin-lsqnonlin-tp4668777p4670940.html

Week 11 and 12: Integrating existing work in optim package.

A recap of the progress in two weeks:

• I had to let go of optimoptions (for GSoC) mainly because of
• time constraints 
• and also because I don't have much experience with objected oriented programming. For optimoptions, I will have to come up with a design. I started with class implementation using classdef as in Matlab, but it is in its infancy in Octave and it could possibly be a limiting factor.
• I am refining my existing functions and including tests and demos so they can be integrated in the optim package.
• lsqnonlin and lsqcurvefit required additional options documentation and OutputFcn and Display setting. These two functions have now been successfully integrated in the optim package. [1]
• Functions nlinfit and quadprog are under review.
• I am working on fmincon now. Still have to discuss which backend should be used. lm_feasible can return Lagrange multipliers, gradient and hessian, but since it adheres to the constraints in all iterations, it behaves differently (from Matlab's algos) and sometimes less efficiently as octave_sqp, which only respects the constraints for the final result.

[1] http://sourceforge.net/p/octave/optim/ci/85d621b7e31a3546383248431b2f340d43edd6da/

Week 10: Preliminary work on optimoptions

Thoroughly checking how optimoptions works in  Matlab.

options = optimoptions (SolverName)

Things to do:

• Identify if Solver name is the right string or function handle

• Cater for multiple Algorithms
   A subset of options for different algorithms.
  
• Transfer relevant options of different solvers to modify/create option.
  oldoptions = optimoptions('fmincon','TolX',1e-10)
  newoptions = optimoptions('lsqnonlin',oldoptions)
  
• Using dot notation or optimoptions to modify previously created options.
  (Second argument in optimoptions can be old options)
  
• Display options:
  Set by user on top (for the current algo)
  Default options
  Options set for other algorithms.

Implementation ideas:

In Matlab, these two calls generate the same options object optim.options.Fmincon:
optimoptions('fmincon')

optim.options.Fmincon
What would be more appropriate, IMO, will be to have a function optimoptions of the following format
opts = function optimoptions(SolverName,varargin)

     ...

    obj = optim.options.SolverName(varargin)

    opts = struct(obj)

    ...

This function will

1. Instantiate the relevant class and request for default options from the solver.
2. Compare the user provided options to add relevant options.
3. Display options of the current algorithm.
4. The output can also be returned in the form of a struct to be compatible with optimget.