MAgPIE - An Open Source land-use modeling framework

4.8.2

created with goxygen 1.4.4

Optimization (80_optimization)

Description

This module takes care of the model optimization of the main model, allowing for switching between optimization procedures. It has been introduced to play with different ways to affect the runtime performance of the model via more optimized model solution strategies. The interfaces to the rest of the model are quite limited as it only requires the variables to be optimized vm_cost_glo (total costs) and vm_landdiff (gross land use changes compared to last time step) as direct input. The latter was introduced to select out of a range of cost optimal patterns that one which is closest to the pattern of the previous time step. While CONOPT returns this solution by default, CPLEX does not.

Interfaces

Input

Realizations

(A) lp_nlp_apr17

In this realization, instead of directly starting the nonlinear optimization, a linear version of the model is solved beforehand. In order to linearize the model all nonlinear terms are fixed to best guesses for the respective values. The linear solution serves as an improved starting point for the nonlinear optimization.

All nonlinear terms are fixed to best guess values via nl_fix.gms files which must be provided for each nonlinear module realization.

$batinclude "./modules/include.gms" nl_fix

After all nonlinearities have been fixed the linear model is solved. Via setting magpie.trylinear = 1 the following solve statement starts a linear optimization if no non-linearities remain in the model (Please note that the solve statement still declares a nonlinear / nlp problem even though we expect it to be linear!). Solve statement is put twice for improved model results, in particular for matching LHS and RHS of equations.

    solve magpie USING nlp MINIMIZING vm_cost_glo;
    if(s80_secondsolve = 1, solve magpie USING nlp MINIMIZING vm_cost_glo; );

A second optimization makes sure that in case of a flat optimum that solution is chosen for which the difference in land changes compared to the previous timestep is minimized. This is achieved by setting the calculated total costs of the previous optimization as upper bound and minimizing the land differences.

    if ((magpie.modelstat=1 or magpie.modelstat = 7),
      vm_cost_glo.up = vm_cost_glo.l;
      solve magpie USING nlp MINIMIZING vm_landdiff;
      if(s80_secondsolve = 1, solve magpie USING nlp MINIMIZING vm_landdiff; );
      vm_cost_glo.up = Inf;
    );

After the linear optimization all nonlinear variables are released again.

$batinclude "./modules/include.gms" nl_release

In case that no feasible solution for the linear model is found the best guess estimates for the fixations of nonlinear terms are slightly relaxed to increase the likelihood of finding a feasible solution and the linear solve is repeated. Such as the nl_fix.gms and nl_release.gms rules also nl_release.gms must be provided by the corresponding module realizations.

    if((p80_modelstat(t) <> 1),
$batinclude "./modules/include.gms" nl_relax
    );

Finally, the linear solution is used as starting point for the nonlinear optimization of the model in its full complexity.

  solve magpie USING nlp MINIMIZING vm_cost_glo;
  if(s80_secondsolve = 1, solve magpie USING nlp MINIMIZING vm_cost_glo; );

Limitations This realization requires that all module realizations with nonlinear terms provide a nl_fix.gms and nl_release.gms which fix and release all nonlinear terms in the module. If this is missing and there are still active, nonlinear terms in the linear solve attempt the model run will be cancelled by an error.

(B) nlp_apr17

In this realization the model is solved directly using nonlinear optimization. If the optimization returns an infeasible solution the solve is repeated, either until a feasible solution is found or the maximum number of iterations as defined in s80_maxiter is reached.

solve magpie USING nlp MINIMIZING vm_cost_glo;

Optional second solve statement

if(s80_secondsolve = 1, solve magpie USING nlp MINIMIZING vm_cost_glo; );

Limitations There are no known limitations.

(C) nlp_par

In this realization the model is solved directly using nonlinear optimization. However, the regions are solved in parallel. This allows to use higher spatial resolution, but works only with fixed trade patterns (exo trade realization). To derive these trade patterns a normal run with global optimization is needed. The start script highres.R illustrates how this can be used. If the optimization returns an infeasible solution the solve is repeated, either until a feasible solution is found or the maximum number of iterations as defined in s80_maxiter is reached.

Limitations There are no known limitations.

Definitions

Objects

Sets

Authors

Jan Philipp Dietrich, Todd Munson

See Also

10_land, 11_costs

References