Getting nursing home residents out in time: a Hamburg evacuation routing benchmark

What is it about?

When a wartime bomb is uncovered or a storm surge threatens a low-lying district, emergency services may have to move hundreds of nursing home residents who cannot walk unaided. Deciding which vehicle goes to which facility, how many people it takes, and which shelter it delivers to is a hard scheduling problem — and a commander needs a workable plan in minutes, not hours. We define this as a distinct routing problem and release three Hamburg test cases built from real care-facility records and real road-network travel times: a bomb exclusion zone in Altona-Altstadt (479 residents, one shelter), a storm surge in Wilhelmsburg (510 residents, three shelters), and a larger synthetic stress test (1,348 residents, five shelters). Unlike standard delivery benchmarks, ours captures four realities of institutional evacuation: facilities too large for a single vehicle, boarding times that grow with the number of passengers assisted, limited shelter space, and vehicles shuttling directly between shelter and facility without returning to a depot. We then compare a rule-based dispatcher against three optimization methods — a genetic algorithm, a memetic algorithm, and adaptive large neighbourhood search — each restricted to 300 seconds on a single processor core. A web-based tool lets planners configure scenarios, run the optimizer, and inspect the resulting schedules on a map and timeline.

Featured Image

Photo by John Cardamone on Unsplash

Photo by John Cardamone on Unsplash

Why is it important?

Vehicle routing is one of the most heavily studied problems in optimization, but the benchmarks the field trains and competes on — Solomon's instances, CVRPLIB, the DIMACS challenge — describe commercial delivery: a depot, a fleet, customers to visit, vehicles that come home at the end. An evacuation works nothing like that. A nursing home with 200 residents cannot be cleared by one vehicle, so multiple vehicles must cooperate and split the pickup. Loading mobility-impaired residents takes time that scales with how many people board, so a full vehicle is slow to fill and slow to unload. Shelters have finite space, so the nearest one is not always available. And vehicles shuttle back and forth between facility and shelter without returning to a depot, chaining trips indefinitely until everyone is out. Each of these features exists somewhere in the literature in isolation; none of the standard testbeds combines them, which means methods are being refined against a problem that omits the constraints determining whether an evacuation actually finishes in time. That omission has consequences. Once these constraints are present and a realistic five-minute planning deadline is enforced, no method wins across the board: adaptive large neighbourhood search produced the fastest evacuations in two of five configurations, the memetic algorithm in three, and in the largest case the difference amounted to 27 minutes of evacuation time. The reason is a trade-off invisible in delivery benchmarks — every extra stop costs fixed setup time, which is worth paying when it stops a small ambulance fleet from becoming the bottleneck and counterproductive when large buses are already nearly full. In two configurations the sophisticated search failed to beat a simple dispatching rule within the budget, so any deployed system should return whichever plan is better. For planners, the practical message is that algorithm choice should follow the fleet composition and the geography of the incident. For researchers, we release scenario definitions, travel-time matrices, code, and random seeds so this gap in the benchmark landscape can be tested against directly.