Compare score-function and direct-gradient policy training

Determine whether the score-function estimator used for one-step actor–critic training provides an advantage over direct differentiation through the mixture reward for training a shared policy that selects and weights manufacturing-domain predictors.

Background

The paper formulates selective transfer across manufacturing domains as a contextual, one-step reinforcement-learning problem. A policy assigns weights to candidate predictors, including individual source experts, source subsets, pooled source data, and target-only learning; its reward is based on validation performance relative to supervised reference models. The implementation uses a score-function estimator with a Dirichlet policy over mixture weights.

The authors note that labelled target validation data also permit direct optimization of mixture weights and that direct differentiation through the mixture reward could train a shared policy. The experiments do not resolve whether the chosen score-function estimator is superior to this alternative, leaving a methodological comparison open.