GSoC

[Ratmir Mugattarov]

Hello Sage team!
I am interested in working on "Enhanced optimization solver interfaces for Sage" idea for GSoC project.

Is here the right place to connect with mentor?

Do I need to contribute to the Sage project before applying to GSoC?

I have been working with linear optimisation via python interface for some time. 
I used PuLP earlier, currently I am using Pyomo for building an optimiser of the flow through the graph in general form, using free solvers cbc and glpk. 
I am familiar with MathProg (subset of AMPL), representing non-linear constraints in linear form and individual solver settings.

I want to learn more about linear and non-linear solvers, python interface libraries, and work on integration solvers with sage

My second interest is legislation, I am ok with going through and interpreting the licensing rules and untangling licence non-compatibility issues if required.

I have read a developer guide as David advised to other students here.
What should I do next to apply for GSoC?
Can I get in touch with potential mentor and discuss further details?

Thank you
Kindest regards
Ratmir

[Matthias Köppe]

This is the right place. 

In https://trac.sagemath.org/ticket/26511 I have collected various facts about solver interfaces.

The next step for the GSoC application is to find a project of suitable scope and describe it in some detail.

Useful for Sage would be several things:

- Providing new backends for our frontend class MixedIntegerLinearProgram. Using PuLP or Pyomo (which you mentioned) would make sense, but my current favorite is CVXPY. It has a very active user/developer community, and the way it gives access to expression trees could be very useful.

- For a more detailed interface to CBC, we would like to use CyLP -- there are several issues (https://github.com/coin-or/CyLP/issues) that need help with. (Requires Cython)

- Currently we have no interface in Sage to nonlinear programming (except for a very rudimentary SDP class). 

- In Sage, we are interested in representing optimization problems with exact rational/algebraic data and to find exact solutions. One idea to represent such problems is to use a standard modeling system (for example CVXPY) and to use symbolic parameters instead of the constants. After solving a numerical approximation of the problem to optimality with the numerical solver, sometimes one may be able to reconstruct an exact solution and an exact optimality certificate.