Presolve #153

This issue tracks the development of presolve.

Presolve in MP tackles the various kinds of model entities including nonlinearities. It is to be further discussed how much "presolve" is different from "reformulation".

Technically, reformulations are special presolvers (bridges aka JuMP) acting from a single constraint. The presolvers / bridges should support pre- & postsolve of various values like solutions, basis statuses, IIS, sensitivity values, etc., see #184, #170.

Tasks

• Inlining (elimination) of defined variables. Option cvt:dvelim.
  Inilining can greatly simplify linear and quadratic models via out-multiplication of terms.
  • Option cvt:dvelim=1: substitute only those used 1x. Using reference counting information from NL file: not complete for nested DVs, they can be reported "multiply used" although only used 1x (in higher-level DVs which are indeed multiply used.)
  • Option cvt:dvelim=2: substitute all algebraic uses. Making default - advantage on some examples.
  • Better, using own reference counting and also for DVs created during reformulations Inline algebraic expressions #266.
• Scaling. Some users have huge and/or very different scale in the constants, and Xpress scaling=8 works best then.
• An option to control out-mulitplication: cvt:multoutcard. Are there cases when it's not best?
• Context propagation: Redefinition context #248.
  • Propagate context into arguments of numberof. Not by default - option cvt:pre:ctx2count Gurobi 12.0.1 different (and much slower) than 12.0.0 #267.
• Unify domains Unify variable/expression domains #268.
• Convert AMPL's || back to ==> if one side compares a binary variable. Not to proceed currently as || seems better for Gurobi 13 and equally good for Knitro 15.1.0 on pooling_setup. Possibly, for Gurobi because indicators are converted to SOS.