Mass evacuations in response to national emergencies are complex undertakings in which a patchwork of local, state, and federal agencies cooperate to move whole communities out of evacuation zones to safety. Significant effort has been spent studying how to motivate the community to evacuate and executing traffic management (pointing people to major roads, using contraflow, etc..), but few resources exist for individually tailored destination selection and route planning.

Our approach uses historical, real-time, and forecast data to inform a simulated environment which is used to train a reinforcement learning algorithm. During an emergency, the algorithm is provided with up-to-date forecasts (e.g. waterway flooding probabilities) and real-time updates (e.g. road closures or gas station outages) and then provides adaptive guidance as the situation on the ground changes. This would address common concerns that research has shown prevents individuals from complying with mandatory evacuation orders and should improve the evacuation process by alleviating burdens on major roads, better utilizing the closest available safe lodging, and distributing refueling needs across a more extensive network of refueling stations.