Significantly improved departure predictions thanks to machine learning

Excellent passenger satisfaction levels at Golden Gate Bridge, Highway & Transportation District

Passengers on San Francisco Bay are now well informed about the exact departure times of their buses on all routes. In a pilot project jointly run by INIT and the public transport provider Golden Gate Bridge, Highway & Transportation District, artificial intelligence and machine learning (ML) software from INIT’s subsidiary, inola significantly improved the reliability of departure predictions. This reliable ML prediction will be a key component in INIT’s MOBILE-ITCS nextGen.

The initial pilot project started in March 2019. Baseline accuracy measurements were taken the same year, in June. The prediction engine was installed in August and the very first prediction data with ML was collected in October 2019. In June 2020, the “live” ML pilot project began on selected lines and since then, all lines (around 200 buses) have been integrated. Golden Gate was an ideal pilot project, because in their region the quality of the departure predictions varied immensely, especially for rural and mixed urban/rural journeys, and the public transport company needed more reliable predictions.

Fully automated machine learning process

The basis for the machine learning prediction is inola’s ML-Core. This software provides predictions using historical data and real-time information. It is independent from platforms and operating systems and can process large amounts of data (Big Data). In MOBILEstatistics, INIT’s system for analysis and statistics, the operational data (e.g. GPS data) is collected and processed with a lot of additional information, as well as historical driving times. Based on this data, various trainer systems are available to the ML-Core. The best one is automatically recognised and used by the software. After each training session using processed historical data, the ML model is updated and therefore improved – so that a suitable model is available at all times.

Based on the trained model, the ML-Core calculates driving time predictions for all trip sections. The ML prediction then compiles the newly predicted departure times for the stops from these individual values and transmits them to various processes. Passenger Information shows the live predictions, taking into account possible current traffic restrictions or the driving time of the previous vehicle.

Clear improvement in prognosis quality

The reliability of the prognoses has increased immensely in San Francisco due to machine learning and the continuous training of the model, which is illustrated by the table above (Tab. 1), comparing the predictions on comparable days on selected lines without ML (June 2019) and with ML (August 2021, both Saturdays): the white rows show the average prediction accuracy of all daily trips of a line (30, 40, 70, 101) with the specified allowed deviation, at different times before the scheduled departure. The pale green lines show the prediction accuracy with the ML-Core, again with the specified allowed deviation. In the most extreme case (line 101, 6 to 10 minutes before departure), the prediction accuracy has increased from 49 to a remarkable 87.47 percent.

Tab. 1: Comparative overview of prediction with and without machine learning. Permitted deviation within the prediction: at 1 – 5 minutes: -1 to 1; at 6 – 10 minutes: -2 to 3; at 10 – 15 minutes: -3 to 4; at 15 – 30 minutes: -4 to 5.

ML prediction clearly superior to linear prediction

This considerably increased accuracy is the result of the innovative ML prediction in contrast to the common linear prediction. The latter is based on the assumption of a constant speed between two stops. Any obstructions, such as road work, accidents on the route, or other events (scheduled at short notice) are not taken into account in a linear prediction. The time of day or different days of the week are also often not taken into account. The new prediction quality is definitely something to be proud of. In INIT’s MOBILE-ITCS nextGen, the ML prediction will be included as standard, which means that our ITCS customers will benefit from it in the future.

Passengers will be able to plan their journeys more easily thanks to ML prediction, which will increase customer satisfaction. In San Francisco, a new era in passenger information has already begun.


Kai Brückner

Managing Director