#include <sddp.h>

Inheritance diagram for gcop::SDdp< T, nx, nu, np >:

Collaboration diagram for gcop::SDdp< T, nx, nu, np >:

Public Member Functions
	SDdp (System< T, nx, nu, np > &sys, Cost< T, nx, nu, np > &cost, vector< double > &ts, vector< T > &xs, vector< Vectorcd > &us, Vectorpd *p=0, bool update=true)
virtual	~SDdp ()
void	Iterate ()
void	Forward ()
void	Backward ()
void	Linearize ()
bool	pd (const Matrixnd &P)
bool	pdX (const MatrixXd &P)
void	Reset ()
Public Attributes
int	N
	number of discrete trajectory segments
double	mu
	current regularization factor mu
double	mu0
	minimum regularization factor mu
double	dmu0
	regularization factor modification step-size
double	mumax
	maximum regularization factor mu
double	a
	step-size
double	current_a
	current step-size
std::vector< Vectorcd >	dus
	computed control change
std::vector< T >	xss
	Sample states (N+1) vector.
std::vector< T >	xsprev
	Mean of Sample states (N+1) vector.
std::vector< Vectorcd >	kus
std::vector< Vectorcd >	du_sigma
	Stores the stdeviation at each segment.
std::vector< Matrixcnd >	Kuxs
std::vector< Vectorcd >	Qud
	Inverse Variance for sampling du.
std::default_random_engine	randgenerator
	Default random engine.
std::normal_distribution< double >	normal_dist
	Creates a normal distribution.
Vectornd	Lx
Matrixnd	Lxx
Vectorcd	Lu
Matrixcd	Luu
Matrixnd	P
Vectornd	v
Vectorcd	duscale
	Scales sampled du according to this vector.
Vectornd	dxscale
	Scales sampled dx0 according to this vector.
RenderFunc *	external_render
	RenderFunction for rendering samples.
double	V
Vector2d	dV
double	s1
	Armijo/Bertsekas step-size control factor s1.
double	s2
	Armijo/Bertsekas step-size control factor s2.
double	b1
	Armijo/Bertsekas step-size control factor b1.
double	b2
	Armijo/Bertsekas step-size control factor b2.
char	type
	type of algorithm (choices are PURE, DDP, LQS), LQS is default. In the current implementation second-order expansion of the dynamics is ignored, so DDP and LQS are identical. Both LQS and DDP are implemented using the true nonlinear dynamics integration in the Forward step, while PURE uses the linearized version in order to match exactly the true Newton step.
int	Ns
	Number of samples for linearizing.
Static Public Attributes
static const char	PURE = 0
	PURE version of algorithm (i.e. stage-wise Newton)
static const char	DDP = 1
	DDP version of algorithm.
static const char	LQS = 2
	Linear-Quadratic Subproblem version of algorithm due to Dreyfus / Dunn&Bertsekas / Pantoja.

template<typename T, int nx = Dynamic, int nu = Dynamic, int np = Dynamic>
class gcop::SDdp< T, nx, nu, np >

Constructor & Destructor Documentation

template<typename T , int nx, int nu, int np>

gcop::SDdp< T, nx, nu, np >::SDdp	(	System< T, nx, nu, np > &	sys,
		Cost< T, nx, nu, np > &	cost,
		vector< double > &	ts,
		vector< T > &	xs,
		vector< Vectorcd > &	us,
		Vectorpd *	p = `0`,
		bool	update = `true`
	)

Create an optimal control problem using a system, a cost, and a trajectory given by a sequence of times, states, and controls. The times ts must be given, the initial state xs[0] must be set, and the controls us will be used as an initial guess for the optimization.

After initialization, every call to Iterate() will optimize the controls us and states xs and modify them accordingly. Problems involving time-optimization will also modify the sequence of times ts.

Parameters:

sys	system
cost	cost
ts	(N+1) sequence of discrete times
xs	(N+1) sequence of discrete states
us	(N) sequence of control inputs
update	whether to update trajectory xs using initial state xs[0] and inputs us. This is necessary only if xs was not already generated from us.

template<typename T , int nx, int nu, int np>

gcop::SDdp< T, nx, nu, np >::~SDdp ( ) [virtual]

Member Function Documentation

template<typename T , int nx, int nu, int np>

void gcop::SDdp< T, nx, nu, np >::Backward ( )

Backward pass

template<typename T , int nx, int nu, int np>

void gcop::SDdp< T, nx, nu, np >::Forward ( )

Forward pass

References gcop::Manifold< Matrix< double, _n, 1 >, _n >::bnd, gcop::Manifold< Matrix< double, _n, 1 >, _n >::lb, and gcop::Manifold< Matrix< double, _n, 1 >, _n >::ub.

template<typename T , int nx, int nu, int np>

void gcop::SDdp< T, nx, nu, np >::Iterate ( ) [virtual]

Perform one DDP iteration. Internally calls:

Backward -> Forward -> Update. The controls us and trajectory xs are updated.

Reimplemented from gcop::Docp< T, nx, nu, np >.

template<typename T , int nx, int nu, int np>

void gcop::SDdp< T, nx, nu, np >::Linearize ( )

Linearize around existing us and xs by collecting samples

References gcop::Manifold< Matrix< double, _n, 1 >, _n >::bnd, gcop::Manifold< Matrix< double, _n, 1 >, _n >::lb, gcop::normal_dist(), and gcop::Manifold< Matrix< double, _n, 1 >, _n >::ub.