| Title |
Retrospective Evaluation of Dwell-Regulation Policies with Decision-Time Information Constraints in Unattended Metros |
| Authors |
전혜위(Hye-Wi Jeon) ; 이동진(Dong-Jin Lee) |
| DOI |
https://doi.org/10.5370/KIEE.2026.75.11.2562 |
| Keywords |
Action Masking; Automatic Train Regulation; Proximal Policy Optimization; Retrospective Policy Evaluation; Unattended Train Operation |
| Abstract |
In unattended metros, dwell-time adjustment rebalances headways without changing movement authority, but policies evaluated on historical records may use information that is unavailable at the decision time. This study proposes a retrospective evaluation procedure that replays automatic train supervision (ATS) departure events and restricts the target-headway and follower-departure inputs to records available at each decision time, while the arrival of the controlled train is reconstructed retrospectively. The procedure was applied to 212 days of Incheon Line 2 records to compare a no-adjustment baseline, proportional rules, and proximal policy optimization (PPO) policies under identical replay conditions, constraint filters, and train-test splits. The restricted inputs referenced no later record, and the no-adjustment replay reproduced the recorded departures. PPO slightly reduced headway error relative to the baseline but increased delay and lowered plan adherence, and no advantage over the rule was established. Cumulative shifts caused early departures that stepwise constraint checks missed; blocking them preserved the headway reduction. The procedure exposes such trade-offs and verification limits before field testing. |