After a disruption, 3RO re-plans station routes by fixing then selectively releasing: it strictly beats the unrestricted MILP on most instances within 180 seconds
Synopsis
When a disruption hits a station area, routing, platforming and timing must be revised within minutes. The 3RO framework first fixes routes predicted to stay stable using route-specific classifiers, releases a few fixings when the reduced model is infeasible, and later releases more around interacting trains to recover alternatives. On 226 real station disruption instances, within a 180-second budget it found strictly better incumbents than the unrestricted MILP in 56.7% of all instances and 77.4% of hard instances.
Interpretation
Real-time station-area rescheduling no longer relies on a one-shot learned fixing of routes; the fixing set stays adjustable during the solve. Earlier learning-based reduction fixed routes predicted stable and handed the restricted model to the optimizer, with no recovery if that model was infeasible or excluded better alternatives. On 226 real station disruption instances, within a 180-second budget 3RO obtained strictly better incumbents than the unrestricted MILP in 56.7% of all instances and 77.4% of hard instances, and was no worse in about 83% of both groups when ties are counted.
The framework splits the online run into reduction, repair, refinement and optimization, and each stage touches only the binary route-selection variables. Reduction uses an offline-trained Scoper that picks a classifier class and decision threshold separately for each nominal route, fixes routes predicted stable at their nominal values, and leaves timing, sequencing and deviation variables to the MILP. Repair uses a global model that minimizes released fixings, releasing 2.86 on average; refinement releases fixings around trains that are cancelled or have positive waiting or dwell extensions, and on hard instances it gave a better terminal incumbent than the 2RO variant without refinement in 71.9% of cases.
Perspective
The results come from a six-hour timetable at an eight-platform station in the German rail network, with switch failures and platform closures as the two disruption classes, and candidate locations covering about 20% of the station infrastructure. The method targets station dispatching settings that need a feasible operating plan within seconds to minutes, and offline training relies on reference solutions from the original MILP solved under a relative optimality-gap tolerance and a one-hour limit.
Route-stability labels come from a single reference solution, so a route used in another optimal or near-optimal plan may be labeled unstable. Release patterns depend on the chosen repair strategy, and different minimum-cardinality release sets retain different fixings and change downstream solution quality. The 3,600-second unrestricted MILP solution is still a finite-time incumbent, so the extended comparison measures time-to-quality rather than distance to global optimality.
