Kernel-based search

This section documents kernel-based search utilities provided by eb-optimization.

Kernel-based search utilities support continuous or adaptive exploration of parameter spaces by leveraging kernel functions to model response surfaces and guide optimization.

eb_optimization.search.kernels

Generic discrete-search kernels for eb-optimization.

This module defines mechanical optimization primitives for selecting an argmin or argmax over a finite candidate set.

Responsibilities

• Iterate over a discrete candidate set
• Evaluate a scalar objective function
• Apply deterministic tie-breaking rules
• Return the selected candidate and its score

Non-responsibilities

• Defining candidate grids (handled by search.grid)
• Computing domain-specific objectives (e.g., cost, HR@τ, utility)
• Inspecting data distributions or residuals
• Returning diagnostics, plots, or policies

Design philosophy

These kernels are intentionally simple, deterministic, and domain-agnostic. They serve as reusable building blocks for higher-level tuning logic (tuning modules), enabling consistent and auditable optimization behavior across the Electric Barometer ecosystem.

argmin_over_candidates(candidates, score_fn, *, tie_break='first')

Select the candidate that minimizes a scalar score.

Parameters:

Name	Type	Description	Default
candidates	Iterable[T]	Iterable of candidate values (e.g., floats, ints, tuples).	required
score_fn	Callable[[T], float]	Function mapping a candidate to a scalar score.	required
tie_break	Literal['first', 'last', 'closest_to_zero']	Deterministic tie-breaking rule: - "first": first candidate with minimal score - "last": last candidate with minimal score - "closest_to_zero": among ties, choose candidate with smallest absolute value	'first'

Returns:

Type	Description
(best_candidate, best_score)

Raises:

Type	Description
ValueError	If candidates is empty or if score_fn returns a non-finite value.

argmax_over_candidates(candidates, score_fn, *, tie_break='first')

Select the candidate that maximizes a scalar score.

This is the argmax analogue of argmin_over_candidates.