Hong KongLogistics and transport2004 – ongoingMTR Corporation
AI engine for night-time engineering works scheduling
Hong Kong's metro gave night-time maintenance planning to an AI scheduler and estimated a 51% efficiency gain
Problem
MTR can only maintain its lines during a non-traffic window of about four hours each night. Hundreds of maintenance and project jobs competed for track access, engineering trains and staff, and a handful of experts checked every request against safety and operating rules by hand.
Approach
From early 2004 MTR worked with City University of Hong Kong to add an AI engine to its web-based Engineering Works and Traffic Information Management System (ETMS). The engine encodes the safety and operating rules, flags conflicts as soon as a request is entered, and builds weekly and quarterly plans with heuristic search and a genetic algorithm. It went into daily use in July 2004; planners still review and approve every plan at a weekly meeting, and urgent repairs can be inserted by hand.
Results
Takeaway
The system checked requests against encoded rules as soon as staff entered them, moving error detection forward from the weekly meeting. All reported figures were estimates by MTR and the developers.
Categories
- Size
- Large enterprise
- Industry
- Logistics and transport
- Business function
- Production and qualityLogistics and supply chain
- Innovation type
- Process innovation
- AI technology
- Optimization
- Adoption
- Co-developed with a partner
- Evidence
- Company's own claim
Sources
Source summaries
City University of Hong Kong and MTR staff won an award for deployed AI applications. They describe the rule base, the heuristic and genetic-algorithm scheduler and the release history, and state that the AI engine has been in daily use since July 2004. Comparing 21 weeks of historical schedules, they report 5 to 10 hard-rule violations per day in human-made plans and none in the engine's plans.
MTR's chief information officer and the system's academic developer estimate that ETMS improved the efficiency of possession planning by at least 51% and report that "no show" and unscheduled track access rates fell from about 30% and 50% to 4.1% and 6.4% by February 2005. They also describe feedback on work requests arriving in seconds instead of up to a week.
A City University of Hong Kong researcher and an MTR engineer explain that nightly works must fit into a non-traffic window of roughly 4 to 5 hours and that the engine allocates resources to maximise the work done within safety and operating rules. They note that the 2004 system was being replaced with a new platform that could also serve other cities.
GIGAZINE reports on the MTR scheduler, citing New Scientist's 2014 reporting. About 200 rules, including local noise limits, were encoded in machine-readable form and generated schedules still need human approval. It cites an MTR estimate of about US$800,000 in annual savings from more efficient allocation of work.
Full details are in the original sources.
