[Coin-standards] INFORMS / directions

[Leonardo B Lopes]

Dear Colleagues,

  INFORMS San Jose would be a great opportunity to advance some of our
technical and business issues. It would be even better if by the time we
got there we had a sketch of a framework for the work ahead. I am going
to draw my own views of what that should be, and hopefully others can
chip in.

  Quick Summary:

  * We are trying to accomplish an unified mechanism for communicating
*instances* (not models) of mathematical programs, be they linear, 
non-linear, stochastic, overconstrained or complementarity programs.

  Points:

  * There are two parallel tracks: creating an XML-based representation,
and formalizing the MPS and SMPS standards.

  * The main benefit of our success is to simplify the connection between
solvers and modeling environments. If we succeed, then solver authors
will only need to write one driver to give all modeling environments
access to it. The same will be true for modeling environment authors.
Also, writing these drivers will be much easier than it is today, due to
libraries that manipulate XML and libraries we will produce on top of them.

  * Ancillary benefits include: connectivity transparency; easier access 
to other mathematical / statistical / database / report generator / 
etc... software, including presolvers or automatic differentiators.

  Deliverables:

  * For the XML-based standard:
   * A DTD and/or schema formally describing what the data
   being converted looks like;
   * A library that makes writing drivers easy, based on
   OSI, or to be added to OSI;
   * A test suite and a certification library of problems;

  * For the MPS/SMPS standard:
   * A formal document specifying the standard;
   * A test suite and a certification library of problems;

  What is and isn't expected of *you*:

  * Share your experience, knowledge, and brainpower. Some of the 
comments made in the list before summer were quite interesting. Concrete
proposals (at least for specific sections) are even better. I have
contributed a SNLP DTD (which I have revised a couple of times). ILOG
folks mentioned something about a CLP DTD.

  * Illustrative examples in your field of expertise, or in your
application area.

  * We DO NOT expect you to produce large amounts of coding. We are all 
too busy to do that. But we have commitments from myself, John Forrest, 
Bob Fourer, and Gus Gassman, that once we have good specifications, we
will have working software.

  What we already accomplished:

  * Rows and Columns must be complex objects, so that they can be used by
CLP programs correctly. I have incorporated that into the most recent
proposal.

  * There needs to be an efficient algebra for manipulating similar
matrices, for SLP problems. That is also incorporated into the proposal.

  What we need to decide / accomplish next:

  * Do we need a sufficiently general approach for including
distribution-based problems? Is the SMPS approach to these problems
sufficient? Would someone like to contribute a design that is compatible
with the current proposal, or modify it in a suitable way?

  * Do we have a CLP proposal? If not, we should draft one.

  * I have not thought about complementarity. How should we represent 
that class of problems? Suggestions? How about MDPs? Distributions 
problems? etc...?

  * Other items?


Thanks,
Leo.
--
=======================================================================
Leonardo B. Lopes                                      leo at iems.nwu.edu
Ph.D. Candidate                                           (847)491-8470
IEMS - Northwestern University             http://www.iems.nwu.edu/~leo