A C++ template library for solving small quadradic programs very quickly. Based on the Eigen Linear algebra template library. The primary design goal is to be small, fast, and simple.
Currently supports the following features:
- Quadradic programs with only equality constraints
- Quadradic programs with only inequality constraints. This is implemented using primal/dual interior point methods (specifically, the Mehrotra predictor/corrector method). The algorithm follows the exposition in this Masters thesis
- Dynamic and fixed array sizes, as well as both float/double support. Fixed size arrays can improve performance for very small problems (e.g., 4 dimensions) by close to an order of magnitude
Things missing/TODO:
- Combined equality/inequality constraints
- Specialized code for box constraints (e.g.,
$l_i \le x_i \le u_i$ ) - Non-convergence/failure testing (currently assumes problem is feasible and convex)
- (Stretch goal) Quadradically constrained quadradic programs (QCQPs)
- Currently, the template specialization seems to not converge. This could just be a sign that the algorithm requires double precision, but this should be investigated.
Keeping with the Eigen header-only philosophy, the only installation needed is to copy eigen-qp.hpp into your include path and then:
#include "eigen-qp.hpp"This also requires a C++11 compiliant compiler, which in the case of GCC and Intel requires the --std=c++11 flag.
The main interface to the code can be accessed through an API that mimics MATLAB's quadprog function:
namespace EigenQP {
template<typename Scalar, int NVars, int NIneq>
void quadprog(Eigen::Matrix<Scalar,NVars,NVars> &Q, Eigen::Matrix<Scalar,NVars,1> &c,
Eigen::Matrix<Scalar,NIneq,NVars> &A, Eigen::Matrix<Scalar,NIneq,1> &b,
Eigen::Matrix<Scalar,NVars,1> &x);
}In this notation, quadprog solves the following inequality constrained quadradic program:
The template parameters NVars and NIneq can be set at compile time to allow the compiler to generate faster code for solving small problems; if set to -1 (or Eigen::Dynamic), arbitrary (real) matrices are allowed (although the user is responsible for assuring compatible dimensions).
For equality constrained problems, the class EigenQP::QPEqSolver can be instantiated. This is provided for completeness, but the implementation is fairly trivial (simply using Eigen to do a block matrix inversion).