Tuning parameters for the native product-density TPE implementation. More...

#include <pineforge/hpo/sampler.hpp>

Public Attributes
std::uint64_t	startup_trials = 10
	Number of completed trials required before fitting Parzen estimators.

std::uint64_t	ei_candidates = 24
	Number of draws from `l(x)` considered for each fitted-model ask.

double	gamma_fraction = 0.10
	Fraction of completed observations assigned to the good estimator.

std::uint64_t	gamma_cap = 25
	Upper bound on observations assigned to the good estimator.

double	prior_weight = 1.0
	Positive smoothing mass for the numeric prior and categorical pseudocounts.

bool	constant_liar = true
	Whether outstanding candidates are included only in the bad estimator.

std::optional< std::uint64_t >	history_switch = std::nullopt
	Optional completed-observation threshold for bounded models; nullopt means never switch.

std::uint64_t	scale_ei_candidates = 8
	Acquisition draws after the history switch; never exceeds ei_candidates.

std::uint64_t	bad_reservoir_size = 448
	Uniform older non-elite reservoir size, in addition to 64 recent observations.

std::uint32_t	max_threads = 0
	Maximum independent-dimension workers; zero selects min(8, available CPUs).

Detailed Description

Tuning parameters for the native product-density TPE implementation.

Definition at line 121 of file sampler.hpp.

Member Data Documentation

◆ bad_reservoir_size

std::uint64_t pineforge::hpo::TpeSamplerConfig::bad_reservoir_size = 448

Uniform older non-elite reservoir size, in addition to 64 recent observations.

Definition at line 156 of file sampler.hpp.

◆ constant_liar

bool pineforge::hpo::TpeSamplerConfig::constant_liar = true

Whether outstanding candidates are included only in the bad estimator.

No objective value is fabricated and pending candidates are not counted as completed. This scale-independent constant-liar policy discourages concurrent asks from proposing the same region.

Definition at line 147 of file sampler.hpp.

◆ ei_candidates

std::uint64_t pineforge::hpo::TpeSamplerConfig::ei_candidates = 24

Number of draws from l(x) considered for each fitted-model ask.

The candidate with the largest independent-density log ratio log(l(x)) - log(g(x)) is selected.

Definition at line 129 of file sampler.hpp.

◆ gamma_cap

std::uint64_t pineforge::hpo::TpeSamplerConfig::gamma_cap = 25

Upper bound on observations assigned to the good estimator.

The good set size is ceil(gamma_fraction * completed), capped here and leaving at least one observation for g(x) whenever possible.

Definition at line 137 of file sampler.hpp.

◆ gamma_fraction

double pineforge::hpo::TpeSamplerConfig::gamma_fraction = 0.10

Fraction of completed observations assigned to the good estimator.

Definition at line 132 of file sampler.hpp.

◆ history_switch

std::optional<std::uint64_t> pineforge::hpo::TpeSamplerConfig::history_switch = std::nullopt

Optional completed-observation threshold for bounded models; nullopt means never switch.

Definition at line 150 of file sampler.hpp.

◆ max_threads

std::uint32_t pineforge::hpo::TpeSamplerConfig::max_threads = 0

Maximum independent-dimension workers; zero selects min(8, available CPUs).

Linux CPU quotas and affinity constrain the default. This never changes suggestions.

Definition at line 160 of file sampler.hpp.

◆ prior_weight

double pineforge::hpo::TpeSamplerConfig::prior_weight = 1.0

Positive smoothing mass for the numeric prior and categorical pseudocounts.

Definition at line 140 of file sampler.hpp.

◆ scale_ei_candidates

std::uint64_t pineforge::hpo::TpeSamplerConfig::scale_ei_candidates = 8

Acquisition draws after the history switch; never exceeds ei_candidates.

Definition at line 153 of file sampler.hpp.

◆ startup_trials

std::uint64_t pineforge::hpo::TpeSamplerConfig::startup_trials = 10

Number of completed trials required before fitting Parzen estimators.

Definition at line 123 of file sampler.hpp.

The documentation for this struct was generated from the following file:

sampler.hpp

Public Attributes