Skip to content
Publications

Dataset · Zenodo · September 22, 2026

Data and code for "Safe Reinforcement Learning for Shift-Level Supervisory Control in Open-Pit Mining: An Ablation Study of Risk-Sensitive and Distributional Components"

Yershat Sapazhanov, Madina Alimanova, Kanat Orazaliyev, Bauyrzhan Berlikozha, Bauyrzhan Yermagambetov

Abstract

Per-episode evaluation records, experiment notebooks and figure sources for the article"Safe Reinforcement Learning for Shift-Level Supervisory Control in Open-Pit Mining:An Ablation Study of Risk-Sensitive and Distributional Components". The deposit contains the V2 mine-supervision simulator, the notebooks for everyexperiment reported in the article, and six CSV files of per-episode evaluation recordsfrom which every table and figure in the article is computed. No value in the article istranscribed by hand. Evaluation protocol, applied identically throughout: 100 episodes per condition,evaluation seed = 1000 + training seed, greedy action selection. Contents:- data/T_main_episodes.csv - six controllers, five seeds, seven evaluation conditions- data/T_alpha_episodes.csv - CVaR risk-level sweep with the safety shield on and off- data/T_ppo_extended.csv - Safe PPO-Lagrangian trained to 1,000,000 environment steps- data/T_stage2.csv - extended-budget runs on further seeds, and CVaR-IQN at eight times its main budget- data/T_curves.csv - intermediate training checkpoints of both learners- data/T_ddqn.csv - scalar dueling-head ablation, five seeds- notebooks/ - five Colab notebooks, one per experiment stage- figures/ - figure sources at the journal column width See README.md inside the archive for the mapping from each file to the tables andfigures it produces, and for the column definitions.

Reinforcement Learning in RoboticsMining Techniques and EconomicsAdaptive Dynamic Programming Controlsafe reinforcement learningopen-pit miningsupervisory control

More by Bauyrzhan Berlikozha, PhD