The Price of a Degree-Hour in Data-Centre Cooling: Auditing Deep Reinforcement Learning against Model Predictive Control - data and analysis release

Abstract

Derived results and analysis code for the article "The Price of a Degree-Hour in Data-Centre Cooling: Auditing Deep Reinforcement Learning against Model Predictive Control", submitted to Energy and AI. The deposit contains every derived quantity behind the article's tables and figures: per-episode held-out evaluation records for all arms, per-run training summaries for all 102 policies including the dual-variable traces, the epsilon-constraint sweep, the five annexes, the computed price census, the breakeven frontier and bootstrap artefacts, the classical controller tuning logs, and the per-block means and fitted exchange rates for two instrumented data centres. It also contains the analysis code that turns those records into the reported numbers. The code runs against the records as deposited: 25 of the 29 scripts execute with no arguments from the archive root, which was verified from a clean extraction. The remaining four state what they need, being a command-line argument, the manuscript source, or the benchmark's container image.