This study aims to reduce the occurrence of rare but catastrophic events–a crucial yet often overlooked issue in evacuation research–through a dynamic evacuation guidance system designed for fire spread scenarios. The distributed algorithms employed in this system operate without a central server, ensuring resilient guidance capabilities even if some components fail. We evaluate the distributed system using an integrated simulation model that includes fire spread, evacuation agents, and evacuation signage models. The distributed evacuation guidance system is calibrated using a stochastic multi-objective optimization method called SMORM, which takes a human risk-averse attitude into account. We quantify risks potentially involved in the problem using Average Value at Risk (AVaR), a widely-used risk measure. The MOEA/D algorithm is then employed to minimize four objective functions: the means and AVaRs of total evacuation time and fire hazards experienced by evacuees during evacuations. Through our analysis, we identify the occurrence of rare but catastrophic events and analyze the conditions under which they occur.