ldaplusplus::optimization::ArmijoLineSearch< ProblemType, ParameterType > Class Template Reference

#include <GradientDescent.hpp>

Inheritance diagram for ldaplusplus::optimization::ArmijoLineSearch< ProblemType, ParameterType >:

Public Types
typedef ParameterType::Scalar	Scalar
Public Types inherited from ldaplusplus::optimization::LineSearch< ProblemType, ParameterType >
typedef ParameterType::Scalar	Scalar

Public Member Functions
	ArmijoLineSearch (Scalar beta=0.001, Scalar tau=0.5)
Scalar	search (const ProblemType &problem, Eigen::Ref< ParameterType > x0, const ParameterType &grad_x0, const ParameterType &direction)

Detailed Description

template<typename ProblemType, typename ParameterType>
class ldaplusplus::optimization::ArmijoLineSearch< ProblemType, ParameterType >

Armijo line search is a simple backtracking line search where the Armijo condition is required.

The Armijo condition is the following if \(p_k\) is the negative direction and \(g_k\) the gradient. In Armijo line search we search the largest \(a_k \in \{\tau^n \mid n \in \{0\} \cup \mathbb{N}\}\) for which the Armijo condition stands.

\[ f(x_k - a_k p_k) \leq f(x_k) - a_k b g_k^T p_k \]

\(f(x_k) - a_k b g_k^T p_k\) is a linear approximation of the function at \(x_k\) (scaled by \(b\)) that we assume to be the upper bound for the decrease.

In all the above \(p_k\) is assumed to be of unit length.

See Wolfe Conditions.

Constructor & Destructor Documentation

template<typename ProblemType , typename ParameterType >

ldaplusplus::optimization::ArmijoLineSearch< ProblemType, ParameterType >::ArmijoLineSearch ( Scalar  beta = 0.001, Scalar  tau = 0.5  )

inline

Parameters

beta	The amount of scaling to do the linear decrease
tau	Defines the set of \(a_k\) to try in the line search

Member Function Documentation

template<typename ProblemType , typename ParameterType >

Scalar ldaplusplus::optimization::ArmijoLineSearch< ProblemType, ParameterType >::search ( const ProblemType &  problem, Eigen::Ref< ParameterType >  x0, const ParameterType &  grad_x0, const ParameterType &  direction  )

inlinevirtual

Search for a good enough function value in the direction given and the function given.

This function changes its parameter x0 which is passed by reference.

Parameters

problem	The function to be minimized
x0	The improved position (passed by reference)
grad_x0	The gradient at the initial x0
direction	The direction of search which can be different than the gradient to account for Newton methods

Returns

The function value at the final x0

Implements ldaplusplus::optimization::LineSearch< ProblemType, ParameterType >.

---

The documentation for this class was generated from the following file:

• include/ldaplusplus/optimization/GradientDescent.hpp