LBFGS
References: Byrd1995
Provides parameters to tune the limited memory BFGS (LBFGS) optimization [Edit on GitHub]
Keywords
Keyword descriptions
- MAX_F_PER_ITER
Type: integer
Default: 20
Usage: MAX_F_PER_ITER {integer}Description: Maximum number of force evaluations per iteration (used for the line search)
- MAX_H_RANK
Type: integer
Default: 5
Usage: MAX_H_RANK {integer}Description: Maximum rank (and consequently size) of the approximate Hessian matrix used by the LBFGS optimizer. Larger values (e.g. 30) will accelerate the convergence behaviour at the cost of a larger memory consumption.
- PRINT_LEVEL
Type: enum
Default: LOW
Usage: PRINT_LEVEL MEDIUMDescription: How much output is written out by the LBFGS algorithm. Currently this is independent of the &GLOBAL/PRINT_LEVEL setting.
Valid values:
SILENTAlmost no outputLOWLittle output about f and |proj g| every iterationMEDIUMQuite some output about details every iterationHIGHLots of output about changes of active set and final xDEBUGEverything is written out, useful for debugging purposes only
- TRUST_RADIUS
Type: real
Default: -5.29177209E-001 [angstrom]
Usage: TRUST_RADIUS {real}Description: Trust radius used in LBFGS. Not completely in depth tested. Negativ values means no trust radius is used.
- WANTED_PROJ_GRADIENT
Type: real
Default: 1.00000000E-016 [internal_cp2k]
Usage: WANTED_PROJ_GRADIENT {real}Description: Convergence criterion (overrides the general ones): Requested norm threshold of the gradient multiplied by the approximate Hessian.
- WANTED_REL_F_ERROR
Type: real
Default: 1.00000000E-016 [internal_cp2k]
Usage: WANTED_REL_F_ERROR {real}Description: Convergence criterion (overrides the general ones): Requested relative error on the objective function of the optimizer (the energy)
- __CONTROL_VAL
Type: integer
Default: -1Description: Hidden parameter that controls the printing behavior of the LBFGS optimizer for advanced debug purposes. This option overrides PRINT_LEVEL setting if explicit.