Grid search

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

Grid search utilities support systematic exploration of discrete parameter spaces when tuning optimization policies and decision thresholds.

eb_optimization.search.grid

Grid construction utilities for optimization search spaces.

This module provides small, deterministic helpers for constructing bounded, interpretable parameter grids used by offline optimization routines.

Responsibilities: - Create numerically stable, reproducible grids for scalar parameters - Enforce positivity and boundary constraints - Standardize grid behavior across tuners

Non-responsibilities: - Evaluating objectives - Selecting optimal parameters - Performing any optimization logic

Design philosophy: This utility favors bounded, discrete search spaces for interpretability, auditability, and deployability of learned policies.

make_float_grid(x_min, x_max, step, decimals=10)

Create a numerically robust 1D grid over a closed interval.

This utility is used throughout optimization to create bounded, interpretable candidate sets for discrete parameter search (e.g., uplift multipliers, thresholds).

The returned grid:

• starts at x_min
• increments by step
• includes x_max (to the extent permitted by floating-point arithmetic)
• is clipped and de-duplicated for numerical stability

Parameters:

Name	Type	Description	Default
x_min	float	Lower bound for the grid (inclusive). Must be strictly positive.	required
x_max	float	Upper bound for the grid (inclusive). Must be greater than or equal to x_min.	required
step	float	Step size between candidates. Must be strictly positive.	required
decimals	int	Rounding precision used to stabilize floats and de-duplicate.	10

Returns:

Type	Description
ndarray	A 1D array of unique grid values in ascending order.

Raises:

Type	Description
ValueError	If step is not strictly positive, if x_min is not strictly positive, or if x_max < x_min.

Notes

This utility ensures reproducible and stable grid construction for parameter tuning and optimization purposes, while favoring discrete, bounded search spaces for interpretability and deployability.