-
Top-Two Thompson Sampling (TT-TS) strategy: New selection strategy based on Russo (2020) that targets best-arm identification rather than regret minimization. Draws two independent posterior samples per selection and explores challenger reagents when there is genuine uncertainty about which is best. Supports asymmetric thermal cycling (heated_scale/cooled_scale) and disagreement-rate adaptive cycling.
-
GMIC criticality-weighted component rotation: RouletteWheelSelection now uses Gaussian Mutual Information Criticality (GMIC) with a divergence gate to direct thermal cycling toward components where exploration is still needed. Replaces the previous IPR-based criticality with a simpler, equally effective metric.
-
Adding support for custom evaluator functions. This will allow users to input pre defined functions from scripts and notebooks as evaluator functions without having to alter source code directly.