Tail-Risk Assessment and Scenario-Rule Extraction in an Agent-Based Tsunami Evacuation Simulation with Stochastic Road Blockages: A Coastal Urban District Case Study in Japan

Agent-based simulation (ABS) is widely used to evaluate tsunami evacuation; however, the results are often reported only as the mean or median values, which can miss catastrophic tail outcomes. This study examined distributional width and tail risk under uncertainty in earthquake-induced road blockages in the Sanbo School District, Sakai City, Japan. We conducted a road-network ABS with 3,559 Monte Carlo runs and recorded the number of people not evacuated within 100 min. The outputs show substantial variability and a second mode corresponding to near total evacuation failure; the 95th and 99th percentiles were close to the total district population. Using a tail-event threshold of L ≥ 36,500 people not evacuated (≈ 97% of residents), the threshold-exceedance probability is 12.45%. To explain these tail events, we fit a classification and regression tree (CART) model using per-run binary blockage logs and extract road-blocking if-then rules summarizing leaf rules using coverage and conditional tail probability. The resulting interpretable rules identify critical road blockage combinations,, particularly those concentrated near railroad crossings that block major alternative routes, that substantially increase tail risk, thereby providing guidance for prioritizing mitigation measures and contingency planning.