Simulating future household adaptation to sea level rise using agent-based modeling and reinforcement learning

This paper presents a novel decision support tool for community planners that simulates household adaptation to the future impacts of sea level rise. Local sea level rise scenarios are combined with tide predictions to determine impacts on building exposure, electric power outages, and increases in travel times. Reinforcement learning is then used to train heterogeneous agents – each representing one household – how to respond to these impacts based on reward functions. The agents perceive (1) the immediate sea level rise impacts at their building and in their neighborhood, (2) the current properties of the building they occupy, and (3) the costs to implement an adaptive action. At each time step, agents can take one of four actions: do nothing, leave, elevate, or install an electric generator. Trained agents are then passed to an agent-based model to simulate household adaptation to sea level rise at the community level. This model can be used to simulate future status quo conditions and various adaptation policies, such as incentive programs that reduce costs to elevate. The model is applied to a coastal testbed community under an intermediate sea level rise scenario for 2025 to 2100. With no policies in place to influence agent behavior, approximately 30 % of the agents in the model take some sort of action by 2100. To validate the model, it is shown that the status quo results are comparable to other agent-based models of household response to future coastal hazards and that the model replicates stylized facts.