Description

This model demonstrates how to implement a generic simple outer approximation algorithm
for convex mixed-binary non-linear programs.

This model uses two represenations of the model instance: the index
model instance (with original names) and a vector/scalar model
instance. The latter allows to deal with the Jacobian and similar
objects in a generic n by m matrix fashion.

The modifications (OA cuts, solution elemination cuts, fixing of
variables, etc) for the subproblems are done by GAMS source rewriting
in the scalar model instance space.

The only custom part that needs to be done by the user is to identify
the non-linear constraints and the binary variables in the original
indexed model.

Contributor: Michael Bussieck, November 2012

  START Problem specific part

Small Model of Type : GAMS

Category : GAMS EMP library

Main file : oa.gms

$Title Outer Approximation for Convex Minimization Problem with Binary Variables (OA,SEQ=95)

$ontext
This model demonstrates how to implement a generic simple outer approximation algorithm
for convex mixed-binary non-linear programs.

This model uses two represenations of the model instance: the index
model instance (with original names) and a vector/scalar model
instance. The latter allows to deal with the Jacobian and similar
objects in a generic n by m matrix fashion.

The modifications (OA cuts, solution elemination cuts, fixing of
variables, etc) for the subproblems are done by GAMS source rewriting
in the scalar model instance space.

The only custom part that needs to be done by the user is to identify
the non-linear constraints and the binary variables in the original
indexed model.

Contributor: Michael Bussieck, November 2012
$offtext

* START Problem specific part
$if not set model $set model fuel
$call gamslib -q %model%
$goto %model%

$label fuel
$set binvar  status_vm
$set nlequ   costfn_em, oileq_em
$set binvarc binvar(j) = sum(status_vm(j,u1),1);
$set nlequc  nlequ(i) = 1$costfn_em(i) + sum(oileq_em(i,u1),1);
$set solve   ucom min cost
$goto cont

$label procsel
$set binvar  y1_vm, y2_vm, y3_vm
$set nlequ   inout2_em, inout3_em
$set binvarc binvar(j) = 1$y1_vm(j) + 1$y2_vm(j) + 1$y3_vm(j);
$set nlequc  nlequ(i) = 1$inout2_em(i) + 1$inout3_em(i);
$set solve   process min pr
$goto cont

$label batchdes
$set binvar  y_vm
$set nlequ   time_em, obj_em
$set binvarc binvar(j) = sum(y_vm(j,u1,u2), 1);
$set nlequc  nlequ(i) = 1$time_em(i) + 1$obj_em(i);
$set solve   batch min cost
$goto cont
* STOP Problem specific part

* Generic OA algorithm
$label cont

$onecho > convert.op2
dumpgdx jacobian.gdx
objvar
$offecho
$onecho > convert.opt
gams
gmsinsert
objvar
dictmap
$offecho
* Assume last solve statement is the MINLP to be solved
$call gams %model% minlp=convert optfile=1 lo=%gams.lo%
$if errorlevel 1 $abort problems running %model%

Set i         scalar equations
    j         scalar variables
    nlequ(i)  scalar non-linear equations
    binvar(j) scalar binary variables;

* Original binary variables and non-linear equation maps
Set %binvar%, %nlequ%;

$gdxin dictmap
$load i j %binvar% %nlequ%
Alias (*,u1,u2,u3);
%binvarc%;
%nlequc%;

Scalar
    sstat solve status
    mstat model status
    obj   objective function;
Variables
    x(j)  scalar variables;
Equations
    e(i)  scalar equations;

file fx /'insert.gms'/; fx.pw=4095;

put fx 'solve m using rminlp min objvar'
     / 'scalar mstat,sstat,obj;'
     / 'mstat=m.modelstat; sstat=m.solvestat; obj=m.objval;'
     / 'execute_unload "rminlp", mstat, sstat, obj;'
     / 'option minlp=convert; m.optfile=2; solve m using minlp min objvar'
     / '$stop';
putclose;
execute 'gams gams lo=%gams.lo% u1=insert';
abort$errorlevel 'problems solving rminlp';
execute_load 'rminlp', mstat, sstat, obj;
if (not (mstat=%ModelStat.Optimal% or mstat=%ModelStat.Locally Optimal%),
  abort 'rminlp not solved to (local) optimality';
);

* Check if integer feasible
execute_load 'jacobian', x;
if (sum(binvar(j), abs(x.l(j)-round(x.l(j))))<1e-6,
   abort.noerror 'rminlp integer feasible';
);

Set iter                OA iterations / i1*i10 /
    oacutpmarg(iter,i)  OA cut positive marginal indicator
    oacut(iter,i)       OA cut
    mipcut(iter)        MIP cut
    mipcut1(iter,j)     MIP cut binary variables with level 1 in cut;
Parameter
    lb                  lower bound
    ub                  upper bound
    A(i,j)              Jacobian
    oacutcoef(iter,i,j) OA cut coefficients
    oacutrhs(iter,i)    OA cut rhs;

lb = obj;
ub = inf;
mipcut(iter) = no; mipcut1(iter,j) = no;
alias(iter,it);

$set fmt 18:15
loop(iter,
  execute_load 'jacobian', A, e, x;
* Build outer approximation cut:
  loop(nlequ(i)$e.m(i),
    oacutcoef(iter,i,j)$A(i,j) = A(i,j);
    oacutrhs(iter,i) = sum(j$A(i,j), A(i,j)*x.l(j)) - e.l(i);
    oacut(iter,i) = yes;
    oacutpmarg(iter,i) = e.m(i)>0;
  );
* Solve MIP
  put fx;
  loop(oacut(it,i),
    put / 'equation oa_', it.tl:0 '_' i.tl:0 '; '
          'variable boa_', it.tl:0 '_' i.tl:0 ' / lo ' e.lo(i) ', up ' e.up(i) ' /;'
        / 'oa_' it.tl:0 '_' i.tl:0 '.. ';
    loop(j$oacutcoef(it,i,j), put$(oacutcoef(it,i,j)>0) '+';
      put oacutcoef(it,i,j):%fmt% '*' j.tl:0);
    if(oacutpmarg(it,i), put ' =g= '; else put ' =l= ');
    put oacutrhs(it,i):%fmt% ' + boa_', it.tl:0 '_' i.tl:0 ';';
  );
  loop(mipcut(it),
    put / 'equation mc_', it.tl:0 ';'
        / 'mc_' it.tl:0 '.. ';
    loop(binvar(j), if (mipcut1(it,j), put '+' j.tl:0; else put '-' j.tl:0));
    put ' =l= ' (sum(mipcut1(it,binvar),1)-1):0:0 ';';
  );
  put    'model mp / all';
  loop(nlequ(i), put '-' i.tl:0); put '/;'
       / 'solve mp using mip min objvar;'
       / 'scalar mstat,sstat,obj;'
       / 'mstat=mp.modelstat; sstat=mp.solvestat; obj=mp.objval;'
       / 'Parameter binX(*);'
  loop(binvar(j), put / 'binX("' j.tl:0 '") = ' j.tl:0 '.l;');
  put  / 'execute_unload "mip", mstat, sstat, obj, binX;'
       / '$stop';
  putclose;
  execute 'gams gams optcr=0 lo=%gams.lo% u1=insert';
  abort$errorlevel 'problems solving mip';
  execute_load 'mip', mstat, sstat, obj, x.l=binX;
  if (not (mstat=%ModelStat.Optimal% or mstat=%ModelStat.Integer Solution%),
    if (mstat=%ModelStat.Infeasible%         or
        mstat=%ModelStat.Integer Infeasible% or
        mstat=%ModelStat.Infeasible No Solution%,
      abort.noerror 'MIP infeasible', lb, ub;
    else
      abort 'mip not solved to optimality or infeasibility');
  );

* lower bound update
  lb = obj;
  abort.noerror$(lb>ub) 'crossover', lb, ub;

* Add MIP cut
  mipcut(iter) = yes;
  mipcut1(iter,binvar(j)) = x.l(j)>0.5;

* Solve fixed NLP
  loop(binvar(j), put fx / j.tl:0 '.fx = ' round(x.l(j)):0:0 ';');
  put  / 'solve m using rminlp min objvar'
       / 'scalar mstat,sstat,obj;'
       / 'mstat=m.modelstat; sstat=m.solvestat; obj=m.objval;'
       / 'execute_unload "rminlp", mstat, sstat, obj;'
       / 'option minlp=convert; m.optfile=2; solve m using minlp min objvar'
       / '$stop';
  putclose;
  execute 'gams gams lo=%gams.lo% u1=insert';
  abort$errorlevel 'problems solving rminlp';
  execute_load 'rminlp', mstat, sstat, obj;
  if (mstat=%ModelStat.Optimal% or mstat=%ModelStat.Locally Optimal%,
    ub = min(ub, obj);
  );

* Convergence
  abort.noerror$((ub - lb) < 1e-4) 'converged', lb, ub;
);
display 'out of iterations', lb, ub;