Optimizing k -Removal AUC for Urban Infrastructure Networks

We present a single-city case study (Seaside, Oregon) that ranks cross-layer infrastructure nodes (transport, power, and water) by directly optimizing early service-loss AUC. Rather than use proxy centralities, we optimize the evaluation metric directly: the normalized area under thek-removal curve (AUC) on a fixed grid (α ∈ {0.10, 0.20, 0.30}). A two-stage, CPU-only pipeline plans a short prefix with Monte Carlo Tree Search (MCTS; shortlistM, rollouts, iterations fixed for fairness) and completes the order with a lazy greedy (CELF) tail using exact bitset coverage. On Seaside (N=1176 infrastructure nodes;B=4679 buildings), our method attains AUC ≈ 0.714 at α=0.20, outperforming Degree (0.231), PageRank (0.239),k-core (0.263), Collective Influence (0.223), approximate Betweenness (0.247; sampling budget), and an attachment-count heuristic (0.528). Robustness evidence—α-sensitivity (0.10/0.20/0.30), dependency-rule variants (single vs. multi-attachment; AND/OR), demand weights, attachment jitter, and missing buildings—shows that the observed advantages are not an artifact of a single modeling choice. We also report a fair-compute comparison using wall-clock time and method-specific budgets. For k ∈ {5, 10}, we certify small-koptima via exact MILP (SciPy milp) and report optimality gaps and solver statuses. Overall, this single-city case study suggests that directly optimizing thek-removal AUC yields a simple, CPU-only, auditable ranking that aligns with building-level service loss; establishing external validity will require multi-city assessment.