cma.sigma_adaptation

module documentation

step-size adaptation classes, currently tightly linked to CMA, because hsig is computed in the base class

Class CMAAdaptSigmaBase step-size adaptation base class, implement hsig (for stalling distribution update) functionality via an isotropic evolution path.
Class CMAAdaptSigmaCSA CSA cumulative step-size adaptation AKA path length control.
Class CMAAdaptSigmaDistanceProportional artificial setting of sigma proportional to ||m||,
Class CMAAdaptSigmaDistanceProportional2 update es.sigma = self.sigma * _norm(es.mean),
Class CMAAdaptSigmaMedianImprovement Compares median fitness to the 27%tile fitness of the previous iteration, see Ait ElHara et al, GECCO 2013.
Class CMAAdaptSigmaNone constant step-size sigma
Class CMAAdaptSigmaTPA two point adaptation for step-size sigma.
Function tpa_abs_z_with_offset0 return z_abs unmodified
Function tpa_abs_z_with_offset1 return z_abs + z_min
Function tpa_abs_z_with_offset2 return absolute z-value z_abs lower bounded by z_min
Variable csa_dampdown_fac Asymmetric damping parameter read by CMAAdaptSigmaCSA` on instantiation and on reset in `.initialize_settings
Variable tpa_dampdown_fac Asymmetric damping factor, read by TPA on instantation and reset. Is 4 in the restricted CMA of 2024, was tentatively 2, 4 mostly does not change the standard performance much
Class _TPAParameters Parameters for CMAAdaptSigmaTPA,
Function _norm Undocumented
Variable _CSA_cs Undocumented
Variable _CSA_cs_max Undocumented
Variable _CSA_cs_sqrt Undocumented
Variable _CSA_dampfac_mueff Damping for large mueff, the default was 2, however 10 would solve issue #231?
Variable _CSA_dampfac_mueff_attenuation_dimension Damping attenuation for small dimension such that _CSA_dampfac_mueff_inner is multiplied by [1/2, 3/4, 7/8,...] when dimension == attentuation_dimension * array([1, 2, 3,...]) and with a smaller factor in smaller dimension bounded from below to keep dampfac >= 1.
Variable _CSA_dampfac_mueff_inner Damping factor for mueff > n >> 1, the default was 1
Variable _CSA_damps Value for CSA damping d_sigma (ds), replaces the mu-dependent default computation and is dynamically multiplied by the 'CSA_dampfac' option value
Variable _CSA_recompute_params_when_masked When True, recompute cs and ds when the evolution path is masked which changes the effective dimension. This will interfere with a possible dynamic choice of cs or ds via PPO.
Variable _inner_threshold Undocumented
Variable _true_inner Undocumented
def tpa_abs_z_with_offset0(z_abs, params):

return z_abs unmodified

def tpa_abs_z_with_offset1(z_abs, params):

return z_abs + z_min

def tpa_abs_z_with_offset2(z_abs, params):

return absolute z-value z_abs lower bounded by z_min

csa_dampdown_fac: int =

Asymmetric damping parameter read by CMAAdaptSigmaCSA` on instantiation and on reset in `.initialize_settings

tpa_dampdown_fac: int =

Asymmetric damping factor, read by TPA on instantation and reset. Is 4 in the restricted CMA of 2024, was tentatively 2, 4 mostly does not change the standard performance much

def _norm(x):

Undocumented

_CSA_cs =

Undocumented

_CSA_cs_max =

Undocumented

_CSA_cs_sqrt: bool =

Undocumented

_CSA_dampfac_mueff: int =

Damping for large mueff, the default was 2, however 10 would solve issue #231?

_CSA_dampfac_mueff_attenuation_dimension: int =

Damping attenuation for small dimension such that _CSA_dampfac_mueff_inner is multiplied by [1/2, 3/4, 7/8,...] when dimension == attentuation_dimension * array([1, 2, 3,...]) and with a smaller factor in smaller dimension bounded from below to keep dampfac >= 1.

_CSA_dampfac_mueff_inner: int =

Damping factor for mueff > n >> 1, the default was 1

_CSA_damps =

Value for CSA damping d_sigma (ds), replaces the mu-dependent default computation and is dynamically multiplied by the 'CSA_dampfac' option value

_CSA_recompute_params_when_masked: bool =

When True, recompute cs and ds when the evolution path is masked which changes the effective dimension. This will interfere with a possible dynamic choice of cs or ds via PPO.

_inner_threshold: int =

Undocumented

_true_inner: bool =

Undocumented