Concepts#

A run of the algorithm is configured along three independent axes: what you constrain, how wide each confidence interval is, and how those intervals are combined into a single bound. These pages cover one axis each, on top of the mathematical background they share.

Mathematical background

Candidate selection and the safety test, delta splitting, interval arithmetic, and the predicted bound that steers the optimizer.

Mathematical Background
What you constrain

The five shipped fairness definitions, how to choose between them, and the base variables you write your own constraints over.

Fairness constraints
How wide each interval is

Hoeffding, empirical Bernstein, betting and Student’s t — what each assumes, and which is tighter under what conditions.

Concentration Inequalities
How intervals combine

The seldonian_type variants, from uniform delta splitting to affine-form compilation.

Algorithm Variants

See also

The two axes are orthogonal: any inequality may be paired with any seldonian_type.