VW::confidence_sequence_utility Namespace Reference

Functions
float	binary_search (float fhat, float delta, float sens, float tol=1e-6f)
void	get_cost_ranges (float delta, VW::LEARNER::learner &base, VW::multi_ex &examples, bool min_only, float min_cb_cost, float max_cb_cost, std::vector< float > &min_costs, std::vector< float > &max_costs, std::vector< VW::action_scores > &ex_as, std::vector< std::vector< VW::cb_class > > &ex_costs)

Variables
constexpr int	BINARY_SEARCH_MAX_ITER = 20

Function Documentation

binary_search()

float VW::confidence_sequence_utility::binary_search ( float  fhat, float  delta, float  sens, float  tol = 1e-6f  )

inline

Binary search to find the largest weight w such that w * (fhat^2 - (fhat - w * sens)^2) <= delta. See Section 7.1 in https://arxiv.org/pdf/1703.01014.pdf.

get_cost_ranges()

void VW::confidence_sequence_utility::get_cost_ranges	(	float	delta,
		VW::LEARNER::learner &	base,
		VW::multi_ex &	examples,
		bool	min_only,
		float	min_cb_cost,
		float	max_cb_cost,
		std::vector< float > &	min_costs,
		std::vector< float > &	max_costs,
		std::vector< VW::action_scores > &	ex_as,
		std::vector< std::vector< VW::cb_class > > &	ex_costs
	)

Compute cost ranges [min_cost, max_cost] for each action using sensitivity analysis. Used by both RegCB and SquareCB reductions.

Parameters

delta	Threshold on empirical loss difference (controls confidence width).
base	The base learner used for sensitivity queries.
examples	The multiline example set (action predictions are backed up and restored).
min_only	If true, only compute minimum costs (used by RegCB optimistic variant).
min_cb_cost	Lower bound on CB costs (e.g., 0).
max_cb_cost	Upper bound on CB costs (e.g., 1).
min_costs	Output vector of minimum cost for each action.
max_costs	Output vector of maximum cost for each action (unchanged if min_only).
ex_as	Scratch buffer for backing up action scores.
ex_costs	Scratch buffer for backing up CB costs.

Variable Documentation

BINARY_SEARCH_MAX_ITER

constexpr int VW::confidence_sequence_utility::BINARY_SEARCH_MAX_ITER = 20

constexpr