MDVRP-AssignmentNet: synthetic multi-depot VRP instance testbed

This record contains the synthetic Multi-Depot Vehicle Routing Problem (MDVRP) instance testbed used in the master's thesis on learning routing-aware customer-to-depot assignments for the MDVRP (KU Leuven, 2026).

Each instance places customers and depots in the unit square [0,1]². The testbed spans ten size classes at a fixed customer-to-depot ratio of 25, from 50 customers / 2 depots up to 500 customers / 20 depots, and covers two spatial regimes in equal proportion: uniform (customers sampled from a fixed "city pool") and clustered (a Gaussian mixture with a dispersed background component). Vehicle capacity is fixed at Q = 25, integer customer demands are drawn uniformly from {1, …, 10}, and depots are selected from a regular 7×7 candidate grid.

Each instance is solved with PyVRP (Hybrid Genetic Search) to a near-stationary, high-quality solution; the customer-to-depot assignment extracted from the solver routes serves as the supervised label. Instances on which the solver does not converge are excluded. Every instance is fully determined by integer seeds (depot, customer, demand, fleet, and solver seeds), so the testbed is exactly reproducible.

The archive provides the generated instances together with their solver solutions (JSON), organized by size class (n{customers}_d{depots}), as well as the Python scripts used to generate and post-process them.