Supplementary Webpage

Each JSON specifies visits with hard common-start windows, deterministic service durations and caregiver eligibility, caregiver shifts, base endpoints, opening costs and capacities, compulsory openings, a directed nominal travel matrix, endpoint exposure zones, and shared-factor uncertainty. A two-task visit requires two distinct caregivers starting together. All entity IDs and matrix indices are zero-based. Headquarters (base 0) is compulsory, and each active caregiver starts and returns at its assigned base.

Divide timing and opening-cost ticks by time_scale to obtain source time units and travel-equivalent costs. A journey receives ceil(rate × nominal travel ticks) on each distinct endpoint zone; a common endpoint zone is counted once. Factor intensities lie in [0, 1] and sum to at most gamma. Unused matrix locations are removed through consistent reindexing; all retained directed durations are unchanged. Original source identifiers remain in the JSON.

Per-run limits are 600 seconds for sizes 10 and 25, 1,800 seconds for sizes 50 and 100, and 3,600 seconds for sizes 200 and 300. Both methods use one solver thread and 128 GiB of memory. At size 10, the matheuristic’s full-workforce mixed-integer linear programming (MILP) call provides a global certificate. At size 25, all twelve Gurobi runs prove optimality and all 120 matheuristic runs attain the corresponding optimum; 102 matheuristic runs provide their own full-model certificate and 18 stop adaptively without their own certificate. At size 50, Gurobi proves the optimum on RBSSRP_050_H02; five of its ten matheuristic runs attain this externally certified optimum, while all ten stop through adaptive stagnation without their own certificate. Adaptive stagnation alone is not an optimality proof. The linked files contain mathematical inputs; reproducing the recorded algorithm runs also requires the research package’s common initialization, seeds, software, and experimental settings.