Skip to content

Latest commit

 

History

History
118 lines (92 loc) · 4.17 KB

README.md

File metadata and controls

118 lines (92 loc) · 4.17 KB
Raw

L-BFGS-B wrapper for C++

lbfgsb_cpp_wrapper is a simple C++ wrapper around the original Fortran L-BGSG-B routine, one of the most widely-used limited-memory quasi-Newton algorithms for bound-constrained optimization. It tries to be compatible with all king of modern vector-like containers, like std::vector, std::array, armadillo vectors or Eigen vectors.

A quick example

A simple example may be found under the examples folder. We give a short overview of the keys steps required to get your code running.

The files required to use the l_bfgs_b wrapper class are located under the include directory (C++ files) and the Lbfgsb.3.0 directory (Fortran files). To specify the problem to be optimized you need to implement a tiny class extending the templated-class problem, which can be found under the include directory.

template<class T>
class quadratic_problem : public problem<T> {
public:
    // here, 2 is the problem input dimension
    quadratic_problem() : problem(2){}

    double operator()(const T& x) {
        return std::pow(x[0] - 0.5, 2) +  std::pow(x[1] - 1, 2);
    }

    void gradient(const T& x, T& gr) {
        gr[0] = 2 * (x[0] - 0.5);
        gr[1] = 2 * (x[1] - 1);
    }
};

Note that this class specifies the dimension of the problem through the constructor, the objective function and its gradient. To specify the box-constraints of the problems we can use the set_lower_bound and set_upper_bound methods inherited from class problem:

 typedef std::array<double,2> my_vector;
 quadratic_problem<my_vector> qp;
 qp.set_lower_bound({-2, -2});
 qp.set_upper_bound({2, 2});

The optimization is then performed using the l_bfgs_b class:

 my_vector initPoint = {2, 3};
 l_bfgs_b<my_vector> solver;
 solver.optimize(qp, initPoint);
 std::cout << "MINIMUM LOCATED AT : (" << 
           initPoint[0] << ", " << initPoint[1] << ")" << 
           std::endl;

To run the code, it is necessary to link the Fortran routines with the C++ files. Fortunately, it is possible to do it in a straightforward manner using cmake.

mkdir build
cd build
cmake ..
make 
../bin/simple_example

A full example

A full example using all kind of vector-like containers is provided in l_bfgs_b_example.cpp. To run the full example, you will need to download and install the latest versions of armadillo and Eigen, since the example makes use of them. Additionally, you will need to set the environment variable EIGEN3_INCLUDE_DIR to wherever you had installed Eigen.

After the installation process, you can compile the example using cmake:

# move to the same build directory as in the previous code chunk ...
cd build
# ... and clean it
rm -rf *
# change the installation path if needed
export EIGEN3_INCLUDE_DIR="/usr/local/include/eigen3/"
cmake -DBUILD_SIMPLE_EX=off -DBUILD_FULL_EX=on ..
make
../bin/full_example

Install it and compile your code

It is possible to use cmake to install the library and the required C++ headers:

cd build
rm -rf *
cmake ..
# Installation requires root permissions
sudo make install

If you don't have root permissions or you want to specify an installation folder you may use:

cd build
rm -rf *
cmake -DCMAKE_INSTALL_PREFIX:PATH=/path/to/folder ..
make install

After the installation, remember to add the shared-library folder to your LD_LIBRARY_PATH environment variable. It is possible now to compile your programs without cmake:

# Compile simple_example.cpp
cd examples
# change the path to your library if needed
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:"/usr/local/lib/lbfgsb_cpp"
g++ simple_example.cpp -std=c++11 -llbfgsb_cpp -L/usr/local/lib/lbfgsb_cpp/ \
    -o simple_example
./simple_example

License

This project is licensed under the terms of the MPL2.0 license.