Bus-holding study cuts simulated passenger time while protecting journey completion

A bus controller can reduce the time passengers accumulate in an evaluation while leaving more of them short of their destination. Yifan Zhang and Liang Zheng examine that failure in a study of multi-line bus holding, then test a bounded change to the controller: a completion-safety reserve for a specific set of lightly loaded tail-bus states.
The authors report a 3.71% reduction in generalized passenger time compared with the same learned controller without the final reserve. Journey completion also improved. These are simulator-to-simulator results for one fixed candidate and its direct parent, rather than evidence from an operating bus fleet.

Learning without experimenting on passengers

Bus holding delays a vehicle at a stop to help control spacing. The delay can ease bunching, but passengers already aboard spend longer in the vehicle. The study accounts for both waiting and in-vehicle passenger time, using asynchronous stop decisions across twelve directional services and 389 scheduled vehicle trips.
The researchers trained with an immutable SUMO dataset containing 3,284,718 transitions from 240 episodes. Eight behavior policies and thirty environment cells supplied the logged experience. They initialized a transit-adapted H2O+ actor through implicit Q-learning, then continued training for 25,000 decision events in a less expensive, event-driven simulator. The target simulator supplied endpoint evaluation rather than online exploration.
That separation introduces a transfer problem: a cheaper simulator can differ in travel times and how passenger events propagate. Shared state and action contracts help make the comparison interpretable, but they do not establish that a trained controller will transfer to a real transport network.

The reserve changes a narrow set of actions

The final intervention acts on low-load tail buses when demand remains. In the specified states, it raises the physical hold to sixty seconds. The reserve-augmented candidate and direct parent share the same frozen learned actor and inherited action transform; the parent omits this final reserve. No controller parameters change during target-SUMO evaluation.
This paired design makes the final reserve the relevant difference. The authors do not attribute the confirmed improvement to a general superiority of H2O+ or to density-ratio correction. Their final training used unit simulator weights after held-out ratio diagnostics failed to generalize.

Fresh episodes support a limited claim

Formal confirmation used 1,350 new SUMO episodes across ten independent environment blocks, three demand levels and five training seeds, with forty-five fixed physical behaviors. Mean generalized passenger time fell from 2,484.86 to 2,392.73 seconds per departed passenger. Mean completion rose from 0.96185 to 0.96650, while unfinished passengers fell from 655.42 to 578.20.
The primary endpoint improved and the completion noninferiority condition passed in all ten paired blocks. The study reports one-sided exact sign-test p-values of 0.0009766 for both conditions. Candidate development used earlier outcomes, but the final comparison used an untouched formal environment.
For transit researchers, the useful lesson is the separation of a time objective from a service-completion constraint. Evaluating those endpoints together exposes a failure that a lower average time alone can conceal. Field reliability and transfer beyond this simulator scenario family remain outside the confirmed result.