mrob::Optimizer Class Reference

#include <optimizer.hpp>

Inheritance diagram for mrob::Optimizer:

Public Types
enum	optimMethod { NEWTON_RAPHSON =0, LEVENBERG_MARQUARDT_SPHER, LEVENBERG_MARQUARDT_ELLIP }

Public Member Functions
	Optimizer (matData_t solutionTolerance=1e-4, matData_t lambda=1e-5)
uint_t	solve (optimMethod method, uint_t max_iters=1e2, double lambda=1e-5)
virtual matData_t	calculate_error ()=0
virtual void	calculate_gradient_hessian ()=0
virtual void	update_state ()=0
virtual void	bookkeep_state ()=0
virtual void	update_state_from_bookkeep ()=0

Protected Member Functions
uint_t	optimize_newton_raphson ()
virtual uint_t	optimize_newton_raphson_one_iteration (bool useLambda=false)=0
uint_t	optimize_levenberg_marquardt ()
virtual matData_t	calculate_model_fidelity (matData_t diff_error)=0

Protected Attributes
optimMethod	optimization_method_ {}
matData_t	solutionTolerance_
uint_t	max_iters_
MatX1	gradient_
MatX1	dx_
matData_t	lambda_

Detailed Description

The class optimizer provides a level of abstraction for solving second order optimization problems of the form:

min C(x)

The current methods implemented are:

• Newton-Raphson (NR), this is a second order method: x' = x - (Hessian)^-1 * Gradient

  Note that Gauss-Newton is an instance of this, but approximating the Hessian by J'*J

• Levenberg-Marquardt: Spherical approximation x' = x - (H + lambda * I)^-1 * gradient
• Levenberg-Marquardt: Elliptical approximation x' = x - (H + lambda * D)^-1 * gradient, where D = diag(H)

given the following requirements:

• C(x): a Cost function

Member Enumeration Documentation

optimMethod

enum mrob::Optimizer::optimMethod

This enums optimization methods available:

• NR: Newton-Raphson (Gauss_newtow is a variant with approximations in the Hessian)
• LM_S: Levenberg Marquardt: Spherical
• LM_E: Levenberg Marquardt: Eliptical

Member Function Documentation

bookkeep_state()

virtual void mrob::Optimizer::bookkeep_state ( )

pure virtual

For Levenberg-Marquard This function bookkeeps the current state values This is in case the optimization step does not improve

Implemented in mrob::PlaneRegistration, and mrob::FGraphSolveDense.

calculate_error()

virtual matData_t mrob::Optimizer::calculate_error ( )

pure virtual

General abstract functions to implement: Calculate error calculates the current error function

Implemented in mrob::PlaneRegistration, and mrob::FGraphSolveDense.

calculate_gradient_hessian()

virtual void mrob::Optimizer::calculate_gradient_hessian ( )

pure virtual

Gradient calculates the gradient and Hessian This function may be called after calculate_error() or not (some cases of LM). For a general purpose it is required that this function is either: 1) self-contained. No assumptions made and recalculated (redundantly) residuals 2) Most of the times will be called after calculate_error. No recalculations are requires except inside the update_bookkeep_state which will have invalid residuals and need update (less prefered option)

Implemented in mrob::PlaneRegistration, and mrob::FGraphSolveDense.

calculate_model_fidelity()

virtual matData_t mrob::Optimizer::calculate_model_fidelity ( matData_t  diff_error )

protectedpure virtual

calculate_model_fidely returns the difference between the quadratized model and the current error, as required for the LM algorithm.

This is a an abstrct class since the hessian matrix might be different for the sparse and the dense case

Implemented in mrob::OptimizerDense.

optimize_levenberg_marquardt()

uint_t Optimizer::optimize_levenberg_marquardt ( )

protected

Levenberg-Marquardt method, inside will distinguish between elliptic and spherical

Input useLambda (default false) builds the NR problem with lambda factor on the diagonal H' = H + lambda * D. where D depends on which LM has been selected (TODO better)

optimize_newton_raphson()

uint_t Optimizer::optimize_newton_raphson ( )

protected

Optimizes according to the Newton-Raphson algorithm (second order method).

Input useLambda (default false) builds the NR problem with lambda factor on the diagonal H' = H + lambda * D. where D depends on which LM has been selected (TODO better)

optimize_newton_raphson_one_iteration()

virtual uint_t mrob::Optimizer::optimize_newton_raphson_one_iteration ( bool  useLambda = false )

protectedpure virtual

One iteration of the RN method

Implemented in mrob::OptimizerDense.

solve()

uint_t Optimizer::solve	(	optimMethod	method,
		uint_t	max_iters = 1e2,
		double	lambda = 1e-5
	)

Optimization call. Input: optmization method from {NR=0, LM_S, LM_E}

• max_iterations
• lambda: initial value of lambda for LM methods output: number of iterations

update_state()

virtual void mrob::Optimizer::update_state ( )

pure virtual

Updates the current solution

Implemented in mrob::PlaneRegistration, and mrob::FGraphSolveDense.

update_state_from_bookkeep()

virtual void mrob::Optimizer::update_state_from_bookkeep ( )

pure virtual

For Levenberg-Marquard Undoes an incorrect update

Implemented in mrob::PlaneRegistration, and mrob::FGraphSolveDense.

---

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

• src/common/mrob/optimizer.hpp
• src/common/optimizer.cpp