Learned Energy-Saving Control in O-RAN: Fixed-Schedule Bounds and a Reproducible PPO/DQN Benchmark — research artifact

Research artifact for the paper "Learned Energy-Saving Control in O-RAN: Fixed-Schedule Bounds and a Reproducible PPO/DQN Benchmark" (M. Fayzan, I. ur Rahman, J. Arshad). PPO and DQN controllers set the transmit power (35/20/5 dBm) of one gNB, five times per 10 s episode, in a packet-level ns-3.48 / 5G-LENA simulation (3 gNBs x 3 UEs). The record contains:- a frozen, pre-registered held-out benchmark (6 traffic/mobility scenarios x 20 seeds x 6 controllers = 720 audited episodes) with paired bootstrap statistics;- multi-seed training runs of PPO and DQN (model seeds 2026, 2027, 2028) and all checkpoints;- an exhaustive enumeration of all 243 fixed power schedules in two scenarios (1,458 episodes), which bounds what any adaptive controller can gain;- the code, simulator programs and UE slot-rebase patch, frozen protocols with SHA-256 pins, and the scripts that regenerate every table and figure in the paper. Raw per-packet trace CSVs (about 28 GB) are not included and are available from the corresponding author ([e-mail]) on request. See README.md for the contents of each file.