Industries / Government, education and community / Government / Public transit
Routes and timetables
This is the work of deciding where buses and trains go, how often they run, and which vehicle and which driver covers each trip.
5
Milestones achieved, from 1990 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1990
Driver schedules by computer
A transit schedule must cover every trip with a vehicle and a driver while obeying the work rules in the drivers' union contract, and service levels change through the day. A 1990 paper in Interfaces, an operations research journal, by authors from the software firm GIRO, Montreal's transit operator (then the STCUM) and the University of Montréal, reports that HASTUS, a set of scheduling programs built over the previous ten years, saved Montreal's operator more than 3,000,000 dollars a year in driver scheduling, about 3% of operating costs, and 1,000,000 dollars a year in vehicle scheduling. By then it was installed in nearly 40 large cities.
Credit
Jean-Yves Blais and his professor Jean-Marc Rousseau at the University of Montréal, who founded GIRO; the STCUM
2004
Seoul rebuilds its bus network
On 1 July 2004 Seoul replaced a tangle of overlapping private routes with a network in four colors: blue trunk lines on main roads, green feeder lines into neighborhoods, yellow circular lines and red intercity lines. The city now set the routes, while private companies kept running the buses. The transportation research site Mobility Lab reports that bus ridership rose nearly 14% within a year and nearly 20% by 2010.
Credit
the Seoul Metropolitan Government under Mayor Lee Myung-bak
2015
Houston redraws its network overnight
In August 2015 Houston's transit agency, METRO, replaced a hub-and-spoke bus network, largely unchanged since the 1970s, with a grid meant to put frequent service within reach of more of the city, and added night and weekend service. By the count of Rice University's Kinder Institute, local bus ridership grew 1.2% over September 2015 to July 2016, and METRO reported Sunday local bus ridership up 34% from June 2015 to June 2016. Total ridership, helped by new light rail lines, rose 7.5% in the first half of 2016, against a goal of 20% after two years.
Credit
METRO; Jarrett Walker, a consultant on the design
2015
Helsinki's bus on demand ends
From October 2012 to December 2015 Helsinki Regional Transport (HSL) ran Kutsuplus, nine-seat minibuses that riders booked online and that software routed and pooled as requests came in. An evaluation by Aalto University researchers of 82,290 trips found an average of 1.27 riders per vehicle and a public subsidy of more than €3 million in 2015, its last year.
LessonMatching riders by computer did not by itself fill the vans, and occupancy stayed near one rider per vehicle.
Credit
HSL; the evaluation by Nils Haglund, Miloš Mladenović and colleagues
2025
A redesign without new riders
On 29 June 2025 WMATA, the transit agency for the Washington, D.C., region, launched its Better Bus Network, a redrawn Metrobus network. WMATA's year-end report says 42% of bus trips then had scheduled waits under 12 minutes, up from 33%, yet Metrobus ridership fell 7% in the year to June 2026, which WMATA put down to the local economy, service changes by regional partners and "lagging uptake of new bus services". Greater Greater Washington, a local advocacy group, counted a 6.8% fall from June to May.
LessonA network that is more frequent on paper does not by itself bring riders, and one year's count mixes the redesign with everything else that changed..pdf).
Credit
WMATA
Watch for
AI agents draft driver schedules
Via, a transit software company, says agencies can upload their labor agreements and their vehicle and driver limits to its new Scheduling and Supply Studio, and "the AI agent will surface an optimized schedule ready to review and roster". It calls the product the first integrated platform of its kind.
Achieved whenAn agency publishes a comparison of an AI-built schedule against one made by its own schedulers, with paid hours and rule breaches counted, or an outside party tests the tool on real agency data.
Claimed by
Via, May 2026
Open
01
Work rules read by machine
A public test of whether AI models can turn the work rules in a bus drivers' labor agreement into scheduling rules a computer can check. Examples are the maximum spread, the time from a driver's first sign-on to the end of the working day, and the point where overtime starts. Collect ten public agreements, have a person write each scheduling rule as code by hand, and score each model's version against it by running both over sample duties built to test the edge cases.
The briefCredit · unclaimed02
Redesigns measured one way
A public table that measures large bus network redesigns the same way: the ridership change on each redesigned network against similar agencies that made no redesign over the same months. Start with US agencies, whose monthly ridership is public, and add others where data allow.
The briefCredit · unclaimed
Open 01 · the brief
Work rules read by machine
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of whether AI models can turn the work rules in a bus drivers' labor agreement into scheduling rules a computer can check. Examples are the maximum spread, the time from a driver's first sign-on to the end of the working day, and the point where overtime starts. Collect ten public agreements, have a person write each scheduling rule as code by hand, and score each model's version against it by running both over sample duties built to test the edge cases.
Why it is still open
Via says its AI agent reads labor agreements to build schedules but has published no test, and this research found no public benchmark of rule extraction from labor agreements. The rules carry money: a University of California, Irvine study that ran a scheduling program over real schedules from five US agencies found that paying the spread premium after 10 hours rather than 12 raised labor cost by 4% to 7%.
Built with what exists
Public agreements, such as the 2025–2027 agreement between Denver's Regional Transportation District and Amalgamated Transit Union Local 1001, which pays overtime for driving "beyond a spread of eleven (11) hours from the time of first assignment"; a current AI model; Python; Google's open-source OR-Tools solver to check duties against each rule.
First step
Download three public agreements, list every rule that limits how a driver's day can be built, and code the rules of one agreement by hand.
Done looks like
Ten agreements, at least 150 hand-coded rules, and a published table, for each model, of the share of rules whose code gives the same pass or fail as the hand-coded version on every sample duty.
Open 02 · the brief
Redesigns measured one way
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table that measures large bus network redesigns the same way: the ridership change on each redesigned network against similar agencies that made no redesign over the same months. Start with US agencies, whose monthly ridership is public, and add others where data allow.
Why it is still open
Each agency reports its own figures over its own window. Houston's local bus ridership rose 1.2% in its first year and Washington's Metrobus ridership fell 7% in its first. TransitCenter, a transit foundation, writes that agencies redesigning their bus networks are "the only ones in the US attracting increased bus ridership" without giving a comparison group, and an independent evaluation of Washington's redesign started on GitHub in 2026 had no results in November 2026.
Built with what exists
The US National Transit Database's monthly ridership data; archived GTFS timetables from the Mobility Database; Python; a web agent to find and save each agency's own ridership reports (see the Note "Deleting 80% to start over").
First step
Pull monthly bus ridership for Houston and five similar agencies from 2013 to 2017 and plot the gap before and after August 2015.
Done looks like
At least ten redesigns, each with a published estimate of the ridership change against a comparison group and its uncertainty range, at least two agencies' own reported figures reproduced within one percentage point, and code public.