Ax
Ax copied to clipboard
ScaledParetoEarlyStoppingStrategy
Summary:
This commit introduces ScaledParetoEarlyStoppingStrategy
an early stopping strategy for multi-objective optimization problems. The strategy consists of scaling the observed objective toward the reference point, implicitly creating a scaled Pareto frontier. If a trial qualifies for early stopping, it is stopped if it does not dominate the scaled Pareto frontier.
Two different variations of the Pareto scaling, based on a constant uniform scaling and based on the percentiles of each objective, are highlighted in the following plot:
{F1455464697}
The following is a schematic comparison to using logical operations to combine single-objective strategies:
{F1381511964}
See this document for details.
Differential Revision: D53369314