Parametrized IPs

IP_Param stands for Parametrized Integer Program, a mathematical program in the following form:

\[\begin{split}\min_y c^Ty + (Cx)^T y \\ \text{s.t.} \quad By \le b \\ \quad \quad y \ge 0 \\ \quad \quad y_i \in \mathbb{Z} \quad \forall i \in I\end{split}\]

Where \(y\) are the decision variables for the program, \(x\) are parameters, and \(I\) is the set of indices of integer variables. ZERO supports Integer Programs with linear constraints and bi-linear objectives (linear in \(y\)). You can find the API information in MathOpt::IP_Param. This class is an inheritor of MathOpt::MP_Param, and we use the same notation of the abstract.

Modeling the problem

Following the previous example on parametrized Quadratic Programs, consider the following parametrized Integer Program:

\[ \begin{align}\begin{aligned}\max_{y_1, y_2} y_1 + 2y_2 - 2y_1x_1 -3y_2x_2\\\text{s.t.} \quad 3y_1+4y_2 &\le 5\\\quad \quad y_1, y_2 &\in \{0,1\}\end{aligned}\end{align} \]

This is a Knapsack Problem in \(y\), with an objective function parametrized in the \(x\) variables. We assume the matrix C is a matrix with:

  • A number of rows corresponding to the number of \(y\) variables

  • A number of columns corresponding to the number of \(x\) parameters

#include "zero.h"
//[...] Your other code here
//The linear objectives in y
arma::vec                _c(2);
//The matrix of interaction coefficients among parameters and variables
arma::mat                _C = arma::zeros(2, 2);
//The constraint matrix
arma::mat                _A = arma::zeros(1, 2);

//The list of indexes of variable that are integer
arma::vec _integers(_c.size());

//All variables are integer here
for (unsigned int i = 0; i < _c.size(); ++i) = i;

//Invert the sign since we have a maximization
_C(0, 0) = 2;
_C(1, 1) = 3;
_c(0) = -1;
_c(1) = -2;

//Knapsack Constraint
_A(0, 0) = 3;
_A(0, 1) = 4;

arma::vec _b = arma::vec{5};

//Convert to sparse objects
arma::sp_mat      _C2  = arma::sp_mat{_C};
arma::sp_mat      _A2  = arma::sp_mat{_A};

//Explicitly add the bounds without having to put additional constraints.
VariableBounds    bnds = {{0, 1}, {0, 1}};
//Gurobi Environment
GRBEnv      test;
MathOpt::IP_Param ipParam(_C2, _A2, _b, _c, _integers, bnds, &test);

Computing solutions

With \((x_1, x_2) = (-1, 0.5)\), we can provide the value of \(f\) and solve it through Gurobi

// Enter the value of x in an arma::vec
arma::vec x(2);
x(0) = -1;
x(1) = 0.5;

// Uses Gurobi to solve the model, returns a unique_ptr to GRBModel
auto FixedModel = ipParam.solveFixed(x,true);