Interpreting the Policy-Gradient Control Evidence Base: Reproducible Evaluation, Stochastic Perturbations, and Model-Free Control

Abstract

The evidence base for policy-gradient control spans Kl-Regularized Stochastic Policy Gradient for Stable Continuous Control in Off-Policy Reinforcement…, while related work on Sustainable Reservoir Operation and Control Using a Deep Reinforcement Learning Policy… and Control Randomisation Approach for Policy Gradient and Application to Reinforcement Learning… broadens the methodological context. The discussion uses reproducible evaluation, stochastic perturbations, and model-free control as analytical lenses. Rather than ranking reported results, it examines which claims remain comparable across tasks, datasets, and operating conditions. By aligning terminology and evidence requirements, the article offers a more defensible basis for future empirical work. The final recommendations focus on traceable data, bounded claims, and evaluation under meaningful operating conditions.