XALMO_OPTIMIZER_TRUSTR

Controls the trust-region optimization of extended ALMOs. Trust radius is varied in the outer loop. Once the trust radius is chosen (and fixed) the model function can be minized using various approaches. Currently, an iterative conjugate-gradient approach is used and controlled by the inner loop [Edit on GitHub]

Keywords

• ALGORITHM

• CONJUGATOR

• EPS_ERROR

• EPS_ERROR_EARLY

• ETA

• INITIAL_TRUST_RADIUS

• MAX_ITER

• MAX_ITER_EARLY

• MAX_ITER_OUTER_LOOP

• MAX_TRUST_RADIUS

• MODEL_GRAD_NORM_RATIO

• PRECONDITIONER

Keyword descriptions

ALGORITHM

Type: enum
Default: CAUCHY
Usage: ALGORITHM CG

Description: Selects an algorithm to solve the fixed-radius subproblem [Edit on GitHub]

Valid values:

• CG Steihaug’s iterative CG algorithm that does not invert model Hessian

• CAUCHY Compute simple Cauchy point

• DOGLEG Dogleg optimizer

CONJUGATOR

Type: enum
Default: HAGER_ZHANG
Usage: CONJUGATOR POLAK_RIBIERE

Description: Various methods to compute step directions in the PCG optimization [Edit on GitHub]

Valid values:

• ZERO Steepest descent

• POLAK_RIBIERE Polak and Ribiere

• FLETCHER_REEVES Fletcher and Reeves

• HESTENES_STIEFEL Hestenes and Stiefel

• FLETCHER Fletcher (Conjugate descent)

• LIU_STOREY Liu and Storey

• DAI_YUAN Dai and Yuan

• HAGER_ZHANG Hager and Zhang

EPS_ERROR

Type: real
Default: 1.00000000E-005
Usage: EPS_ERROR 1.E-6

Description: Target value of the MAX norm of the error [Edit on GitHub]

EPS_ERROR_EARLY

Type: real
Default: -1.00000000E+000
Usage: EPS_ERROR_EARLY 1.E-2

Description: Target value of the MAX norm of the error for truncated SCF (e.g. Langevin-corrected MD). Negative values mean that this keyword is not used. [Edit on GitHub]

ETA

Type: real
Default: 2.50000000E-001
Usage: ETA 0.1

Description: Must be between 0.0 and 0.25. Rho value below which the optimization of the model function is not accepted and the optimization is restarted from the same point but decreased trust radius. Rho is the ratio of the actual over predicted change in the objective function [Edit on GitHub]

INITIAL_TRUST_RADIUS

Type: real
Default: 1.00000000E-001
Usage: INITIAL_TRUST_RADIUS 0.1

Description: Initial trust radius [Edit on GitHub]

MAX_ITER

Type: integer
Default: 20
Usage: MAX_ITER 100

Description: Maximum number of iterations [Edit on GitHub]

MAX_ITER_EARLY

Type: integer
Default: -1
Usage: MAX_ITER_EARLY 5

Description: Maximum number of iterations for truncated SCF (e.g. Langevin-corrected MD). Negative values mean that this keyword is not used. [Edit on GitHub]

MAX_ITER_OUTER_LOOP

Type: integer
Default: 0
Usage: MAX_ITER_OUTER_LOOP 10

Description: Maximum number of iterations in the outer loop. Use the outer loop to update the preconditioner and reset the conjugator. This can speed up convergence significantly. [Edit on GitHub]

MAX_TRUST_RADIUS

Type: real
Default: 2.00000000E+000
Usage: MAX_TRUST_RADIUS 1.0

Description: Maximum allowed trust radius [Edit on GitHub]

MODEL_GRAD_NORM_RATIO

Type: real
Default: 1.00000000E-002
Usage: MODEL_GRAD_NORM_RATIO 1.E-2

Description: Stop the fixed-trust-radius (inner) loop optimization once the ratio of the current norm of the model gradient over the initial norm drops below this threshold [Edit on GitHub]

PRECONDITIONER

Type: enum
Default: DEFAULT
Usage: PRECONDITIONER DOMAIN

Description: Select a preconditioner for the conjugate gradient optimization [Edit on GitHub]

Valid values:

• NONE Do not use preconditioner

• DEFAULT Same as DOMAIN preconditioner

• DOMAIN Invert preconditioner domain-by-domain. The main component of the linear scaling algorithm

• FULL Solve linear equations step=-H.grad on the entire space