Researchers from Johns Hopkins University and the University of Florida have been awarded a total of nearly $1.2 million by the National Science Foundation’s Fire Science Innovations program to advance AI-enabled tools that predict human behavior during wildfire evacuations. Hopkins principal investigator Susu Xu, an assistant professor of civil and systems engineering, says the project aims to strengthen community resilience by understanding how civilians, incident response teams, and public safety officials make protective action decisions during deadly blazes.
The award allows the team to expand on their prior research that simulates and predicts civilian decision-making during wildfires by using an AI-based model known as FLARE. With support from NSF, the researchers are working to create a system that integrates decision-making among civilians, incident commanders, and safety personnel that accurately predicts behavior to enhance safety and evacuation planning.
FLARE was detailed in “From Perceptions to Decisions: Wildfire Evacuation Decision Prediction with Behavioral Theory-informed LLMs,” published in ACL Anthology in July.
The project will produce simulation methods to promote teaching, training and learning, and support wildfire resilience by allowing public safety officials to use open-access tools.
“During the Los Angeles wildfires last year, we saw a lot of negative effects related to evacuation,” Xu says. “While some people left prior to the fires, others chose to evacuate at the last minute, causing traffic congestion and prompting drivers to abandon their cars on the same roads used by emergency response teams. We also saw that many people did not have the means to evacuate. By using simulation models that predict how groups make decisions, we can support more effective planning and evacuation to reduce those negative effects.”
Current behavior prediction models include behavioral and psychological theory but don’t fully represent the complexity of human decision-making. The team found that by introducing behavioral theory and psychological theory to guide the reasoning process of large language models, or LLMs, they could significantly improve mimicry of the human reasoning process, allowing them to better understand how implicit mental states are shaped under stressful environment, how people are making decisions and how choices vary among individuals and groups.
While AI is known to “hallucinate,” by generating unreasonable or nonsensical thoughts, Xu’s team incorporates behavioral theory and psychological theory and guides the LLMs using survey data on how people respond during wildfires so that the prediction model is expressive with its chain of thought but also grounded in evidence.
“When people are faced with fire, their decision-making can differ drastically based on multiple factors, like their personal background, mental state, transportation options, and shelter accessibility. It’s imperative that we understand their risk and threat perceptions which lead to their decision-making in the event of a wildfire,” says Xu.
The researchers say that their existing FLARE model to understand civilian behavior is transferable across multiple wildfire events, despite differences in state policies and geography.
As the next step in their research, the team plans to model incident commanders’ decision-making and integrate it with the FLARE model to understand how their decisions will impact communities and individuals.
“The goal of the research is not to have people evacuate but to provide tools to help predict which types of people prefer to wait, evacuate at the last minute, or decide not to evacuate at all,” Xu says. “We want to provide data-based guidance to tailor interventions for different populations to prevent scenarios where evacuees become trapped and to ensure that responders aren’t impeded.”
Additional research collaborators include Johns Hopkins civil and systems engineering PhD student Ruxiao Chen, Johns Hopkins Bloomberg School of Public Health associate Chenguang Wang, and University of Florida associate professor Xilei Zhao and PhD student Yuran Sun.