autora-experimentalist-divergence 0.0.2

sampling based on divergence

AutoRA Divergence Experimentalist

The divergence experimentalist identifies experimental conditions $\vec{x}' \in X'$ with respect the distance between existing experimental data $\vec{x}, $\vec{y} and data predicted by a model $\vec{x_pool}, $\vec{y_pred}:

$$ \underset{\vec{x}'}{\arg\max}~sum(d((\vec{x}, \vec{y}), (\vec{x_pool}, \vec{y_pred})) $$

The aim of this experimentalist is to combine novelty and uncertainty by using a distance that combines both: The distance between existing conditions to new conditions and the distance of existing observations to predictions.