AI / ML · Data science & classical MLOther industries (rail transport (train traffic control))
At the Korean Society for Railway autumn conference on October 29, Lee Dong-seop of Daeati presented the AI train-conflict prediction and resolution system planned for the Second Railway Traffic Control Center (about KRW 250 billion, completion targeted for 2028).
Source [1] 전기신문 · 30 Oct 2026 · “(2026 철도학회 추계학술대회) 철도 관제, 경험 의존 관제서 AI 기반 경합 예측·의사결정 지원으로 전환” ↗
Adoption stage: Research — Paper, lab demo or benchmark — no named operator using it.
Summary
At the Korean Society for Railway autumn conference on October 29, Lee Dong-seop of Daeati presented the AI train-conflict prediction and resolution system planned for the Second Railway Traffic Control Center (about KRW 250 billion, completion targeted for 2028). It couples physics-based micro-simulation (gradient, curves, track length, vehicle specs) with AI trained on actual operating records and track occupancy, aiming to predict about 2 hours ahead and compute conflicts and resolution options in about 3 seconds; resolution runs reinforcement learning first, which handles about 995 of 1,000 cases, with a CP-SAT solver for the rest. The system presents conflict types and options and the dispatcher makes the final choice on punctuality, cost and energy; the Guro control center is to be upgraded to the same system for redundancy.
Kenny's take
What changes in the control room is the job: from spotting conflicts by eye to choosing among resolution options the AI raises two hours ahead. The method is the notable part — not an LLM but physics simulation plus reinforcement learning, with a CP-SAT solver closing the 5 in 1,000 cases the policy cannot. The presenter was explicit that the dispatcher makes the final choice. Stage is research. It moves to pilot when results on live operating data from a named line appear before the 2028 build; if 995 of 1,000 stays a simulation number, this remains a design presentation.
AI draft · approved by Kenny
Analysis
Stage: research. The source is a conference presentation of a system for a control center targeted for 2028; it names the contractor and the project but does not state that the system is running on live train operations or in a bounded trial, and Lee lists "validation across all lines" as future work. The 995-of-1,000 figure is presented as the method's result, without the source saying whether it came from live or simulated cases.
For the pattern hybrid-rl-solver-decision-support and q-ds-vs-llm, the method is the point: no LLM is mentioned; the stack is physics simulation, a learned correction from operating records, reinforcement learning for most conflicts and a CP-SAT constraint solver as the fallback for the roughly 5 in 1,000 the policy cannot resolve.
For q-agent-ops-hitl, the human gate is explicit: the system proposes conflict type and resolution options, and the dispatcher selects the final option on punctuality, cost and energy, with Lee stating the aim is to inform the dispatcher's adjustment rather than replace the dispatcher's judgement.
Facts from the source
- F1On October 29 at the 2026 Korean Society for Railway autumn conference (Phoenix Island, Jeju), Lee Dong-seop, executive director (상무) at Daeati, presented "AI-based micro-simulation for train conflict prediction and resolution". [1]
- F2The Second Railway Traffic Control Center has a total project cost of about KRW 250 billion and targets completion in 2028; it will introduce an AI and big-data-based train management system (TMS) and upgrade the Guro control center to the same system so the two centers are redundant. [1]
- F3Current control depends on dispatchers' experience to adjust train plans and handle conflicts; when several trains become entangled it is hard to identify conflicts and decide the right action. [1]
- F4The system combines physics-based micro-simulation (gradient, curves, track length, vehicle specifications) with AI that learns from actual operating records and track occupancy times; it detects conflicts from predicted train positions and produces resolution options for the situation. [1]
- F5Conflict resolution applies reinforcement learning first and a CP-SAT solver for cases it cannot solve: "reinforcement learning resolves about 995 of 1,000 cases and the solver handles the rest mathematically"; the target is to predict operations about 2 hours ahead with computation at about 3 seconds. [1]
- F6The system presents the conflict type and resolution options and the dispatcher selects the final option considering punctuality, cost and energy; Lee said "the goal is to provide the information needed for operational adjustment quickly rather than have AI replace the dispatcher's judgement", with weather data, algorithm upgrades and validation across all lines to follow. [1]
Organisations: Daeati (대아티아이), Second Railway Traffic Control Center (제2철도교통관제센터)
Counterpoint — verification notes
Codex checked each sentence against the source and searched for counter-evidence. Verdict: Codex check · partly supported.
- Other sideReproduces Daeati’s September 30, 2026 amended filing, which moves its control-system contract end from June 2028 to June 2029. This raises a schedule question but does not by itself disprove the article’s stated 2028 target for the center. [2]
2 sentence-level checks
- summary_kopartly-supported: 약 3초 안에 경합과 해소안을 산출하며 — The source presents roughly three-second computation as a target. The Korean summary phrases it as a capability and changes ‘roughly three seconds’ to ‘within roughly three seconds’; the English summary retains the target qualifier.
- analysispartly-supported: The source names the contractor. — The article names a Daeati executive as presenter but does not identify Daeati as the contractor. A separate railway authority release confirms that role.
Sources
- [1] 전기신문 — “(2026 철도학회 추계학술대회) 철도 관제, 경험 의존 관제서 AI 기반 경합 예측·의사결정 지원으로 전환” original
- [2] dartbrief.com counter-evidence (Codex)
Keep reading
Stay in the loop.