Struct PANOCCache

Source

pub struct PANOCCache<T = f64>where
    T: Float + LbfgsPrecision + Sum<T>,
{ /* private fields */ }

Expand description

Cache for PANOC

This struct carries all the information needed at every step of the algorithm.

An instance of PANOCCache needs to be allocated once and a (mutable) reference to it should be passed to instances of PANOCOPtimizer

Subsequently, a PANOCEngine is used to construct an instance of PANOCAlgorithm

Implementations§

Source §

impl<T> PANOCCache<T>
where T: Float + LbfgsPrecision + Sum<T>,

Source

pub fn new( problem_size: usize, tolerance: T, lbfgs_memory_size: usize, ) -> PANOCCache<T>

Construct a new instance of PANOCCache

§Arguments

problem_size dimension of the decision variables of the optimization problem
tolerance specified tolerance
lbfgs_memory_size memory of the LBFGS buffer

§Panics

The method will panic if

the specified tolerance is not positive
memory allocation fails (memory capacity overflow)

§Memory allocation

This constructor allocated memory using vec!.

It allocates a total of 8*problem_size + 2*lbfgs_memory_size*problem_size + 2*lbfgs_memory_size + 11 floats of type T

Source

pub fn set_akkt_tolerance(&mut self, akkt_tolerance: T)

Sets the AKKT-specific tolerance and activates the corresponding termination criterion

§Arguments

akkt_tolerance: Tolerance for the AKKT-specific termination condition

§Panics

The method panics if akkt_tolerance is nonpositive

Source

pub fn cache_previous_gradient(&mut self)

Copies the value of the current cost gradient to gradient_u_previous, which stores the previous gradient vector

Source

pub fn exit_condition(&self) -> bool

Returns true iff all termination conditions are satisfied

It checks whether:

the FPR condition, gamma*||fpr|| < epsilon ,
(if activated) the AKKT condition ||gamma*fpr + (df - df_prev)|| < eps_akkt are satisfied.

Source

pub fn reset(&mut self)

Resets the cache to its initial virgin state.

In particular,

Resets/empties the LBFGS buffer
Sets tau = 1.0
Sets the iteration count to 0
Sets the internal variables lhs_ls, rhs_ls, lipschitz_constant, sigma, cost_value and gamma to 0.0

Source

pub fn with_cbfgs_parameters(self, alpha: T, epsilon: T, sy_epsilon: T) -> Self

Sets the CBFGS parameters alpha and epsilon

Read more in: D.-H. Li and M. Fukushima, “On the global convergence of the BFGS method for nonconvex unconstrained optimization problems,” vol. 11, no. 4, pp. 1054–1064, jan 2001.