1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
|
// CppNumericalSolver
#include <iostream>
#include <Eigen/LU>
#include "isolver.h"
#include "../linesearch/morethuente.h"
#ifndef BFGSSOLVER_H_
#define BFGSSOLVER_H_
namespace cppoptlib {
template<typename ProblemType>
class BfgsSolver : public ISolver<ProblemType, 1> {
public:
using Superclass = ISolver<ProblemType, 1>;
using typename Superclass::Scalar;
using typename Superclass::TVector;
using typename Superclass::THessian;
void minimize(ProblemType &objFunc, TVector & x0) {
const size_t DIM = x0.rows();
THessian H = THessian::Identity(DIM, DIM);
TVector grad(DIM);
TVector x_old = x0;
this->m_current.reset();
do {
objFunc.gradient(x0, grad);
TVector searchDir = -1 * H * grad;
// check "positive definite"
Scalar phi = grad.dot(searchDir);
// positive definit ?
if (phi > 0) {
// no, we reset the hessian approximation
H = THessian::Identity(DIM, DIM);
searchDir = -1 * grad;
}
const Scalar rate = MoreThuente<ProblemType, 1>::linesearch(x0, searchDir, objFunc) ;
x0 = x0 + rate * searchDir;
TVector grad_old = grad;
objFunc.gradient(x0, grad);
TVector s = rate * searchDir;
TVector y = grad - grad_old;
const Scalar rho = 1.0 / y.dot(s);
H = H - rho * (s * (y.transpose() * H) + (H * y) * s.transpose()) + rho * rho * (y.dot(H * y) + 1.0 / rho)
* (s * s.transpose());
// std::cout << "iter: "<<iter<< " f = " << objFunc.value(x0) << " ||g||_inf "<<gradNorm << std::endl;
if( (x_old-x0).template lpNorm<Eigen::Infinity>() < 1e-7 )
break;
x_old = x0;
++this->m_current.iterations;
this->m_current.gradNorm = grad.template lpNorm<Eigen::Infinity>();
this->m_status = checkConvergence(this->m_stop, this->m_current);
} while (objFunc.callback(this->m_current, x0) && (this->m_status == Status::Continue));
}
};
}
/* namespace cppoptlib */
#endif /* BFGSSOLVER_H_ */
|
