StackFeat-RL

StackFeat-RL is presented as computationally much cheaper than base StackFeat while allowing independent episodes to be parallelized. StackFeat-RL extends StackFeat by using a REINFORCE policy to learn feature-retention behavior and optional per-gene penalty modulation. StackFeat-RL selects features through the intersection of top genes by accumulated coefficient magnitude and accumulated selection count. StackFeat-RL uses one ElasticNetCV-selected alpha per outer training fold and holds it fixed during later selection episodes. StackFeat-RL avoids manual specification of panel size, regularisation strength, and stopping criterion.