9
1
mirror of https://github.com/QuantumPackage/qp2.git synced 2024-12-11 06:03:30 +01:00
qp2/src/utils_trust_region/EZFIO.cfg

90 lines
3.5 KiB
INI
Raw Normal View History

2023-02-08 16:44:40 +01:00
[thresh_delta]
type: double precision
doc: Threshold to stop the optimization if the radius of the trust region delta < thresh_delta
interface: ezfio,provider,ocaml
default: 1.e-10
[thresh_rho]
type: double precision
doc: Threshold for the step acceptance in the trust region algorithm, if (rho .geq. thresh_rho) the step is accepted, else the step is cancelled and a smaller step is tried until (rho .geq. thresh_rho)
interface: ezfio,provider,ocaml
default: 0.1
[thresh_eig]
type: double precision
doc: Threshold to consider when an eigenvalue is 0 in the trust region algorithm
interface: ezfio,provider,ocaml
default: 1.e-12
[thresh_model]
type: double precision
doc: If if ABS(criterion - criterion_model) < thresh_model, the program exit the trust region algorithm
interface: ezfio,provider,ocaml
default: 1.e-12
[absolute_eig]
type: logical
doc: If True, the algorithm replace the eigenvalues of the hessian by their absolute value to compute the step (in the trust region)
interface: ezfio,provider,ocaml
default: false
[thresh_wtg]
type: double precision
doc: Threshold in the trust region algorithm to considere when the dot product of the eigenvector W by the gradient v_grad is equal to 0. Must be smaller than thresh_eig by several order of magnitude to avoid numerical problem. If the research of the optimal lambda cannot reach the condition (||x|| .eq. delta) because (||x|| .lt. delta), the reason might be that thresh_wtg is too big or/and thresh_eig is too small
interface: ezfio,provider,ocaml
default: 1.e-6
[thresh_wtg2]
type: double precision
doc: Threshold in the trust region algorithm to considere when the dot product of the eigenvector W by the gradient v_grad is 0 in the case of avoid_saddle .eq. true. There is no particular reason to put a different value that thresh_wtg, but it can be useful one day
interface: ezfio,provider,ocaml
default: 1.e-6
[avoid_saddle]
type: logical
doc: Test to avoid saddle point, active if true
interface: ezfio,provider,ocaml
default: false
[version_avoid_saddle]
type: integer
doc: cf. trust region, not stable
interface: ezfio,provider,ocaml
default: 3
[thresh_rho_2]
type: double precision
doc: Threshold for the step acceptance for the research of lambda in the trust region algorithm, if (rho_2 .geq. thresh_rho_2) the step is accepted, else the step is rejected
interface: ezfio,provider,ocaml
default: 0.1
[thresh_cc]
type: double precision
doc: Threshold to stop the research of the optimal lambda in the trust region algorithm when (dabs(1d0-||x||^2/delta^2) < thresh_cc)
interface: ezfio,provider,ocaml
default: 1.e-6
[thresh_model_2]
type: double precision
doc: if (ABS(criterion - criterion_model) < thresh_model_2), i.e., the difference between the actual criterion and the predicted next criterion, during the research of the optimal lambda in the trust region algorithm it prints a warning
interface: ezfio,provider,ocaml
default: 1.e-12
[version_lambda_search]
type: integer
doc: Research of the optimal lambda in the trust region algorithm to constrain the norm of the step by solving: 1 -> ||x||^2 - delta^2 .eq. 0, 2 -> 1/||x||^2 - 1/delta^2 .eq. 0
interface: ezfio,provider,ocaml
default: 2
[nb_it_max_lambda]
type: integer
doc: Maximal number of iterations for the research of the optimal lambda in the trust region algorithm
interface: ezfio,provider,ocaml
default: 100
[nb_it_max_pre_search]
type: integer
doc: Maximal number of iterations for the pre-research of the optimal lambda in the trust region algorithm
interface: ezfio,provider,ocaml
default: 40