AN AGENT-BASED MODEL OF HURRICANE EVACUATION DECISIONS Seth McGinnis Institute for the Study of Society & Environment National Center for Atmospheric Research Boulder, CO. 80307-3000 ABSTRACT Despite its importance, the social component of evacuation is usually neglected in planning and modeling hurricane response. We propose to develop an agent-based model of a threatened population’s collective evacuation decision process that takes into account differential response due to demographics, source and flow of information, and social network effects. In addition to improving our understanding of the evacuation problem, this model will serve as an educational tool to aid decision-makers and will build quantitative modeling capability in ISSE. THE PROBLEM OF HURRICANE EVACUATION Hurricanes Katrina and Rita demonstrated the tremendous hazard hurricanes present to coastal communities, highlighting evacuation vulnerabilities. Despite the evacuation of more than one million people from the area, more than 1100 people died when Katrina struck New Orleans because traditional evacuation planning and policy failed to prepare for differential evacuation behavior due to social, economic, and psychological reasons. Similarly, when hurricane Rita moved towards the Houston/Galveston area, existing evacuation plans were unable to cope with a massive shadow evacuation spurred by the psychological effects of Katrina, and thousands of evacuees who would have been safer at home were stranded on highways as the storm approached. Existing hurricane evacuation models fail to account for social factors that dynamically interact to influence individual decision making and drive evacuation behavior. Past surveys of evacuations and current evacuation modeling tools are inadequate for studying decision dynamics, the influence of urban design, and the importance of demographic variables. This poorly understood individual decision-making process dominates the effectiveness of evacuations and adequacy of public sector response/support. Much of the past work on evacuation demand has taken an engineering or economic (utility-theory) approach; we believe that social-scientific theory supports novel, quantitative exploration of evacuation decision dynamics. Consensus regarding the importance of quantitative social and behavioral modeling in evacuation research is growing among key research advisory boards and organizations, including the National Science Board, NOAA’s Science Advisory Board, the Transportation Research Board, the Office of the Federal Coordinator for Meteorology, and the American Geophysical Union. Nevertheless, the currently most sophisticated crisis-behavioral models are simply sequential logit models1 calibrated to predict demand for hurricane-induced evacuation of cities. Although they can fit observed evacuations using few parameters and there is evidence that models calibrated for one hurricane are applicable to others, these models lack responsiveness to social networks, public policy, and demographic/geographic detail; they do not distinguish between the decision to evacuate and implementation of that decision. Furthermore, they do not incorporate information flow, feedback loops, or causality in decision-making. The social aspect of hurricane evacuation also affects the utility of forecast information: predicted storm track accuracy is immaterial if the public's evacuation decision does not take it into account. Our project is an effort to understand this human element so that it can be incorporated into evacuation policy, planning, research, and modeling. We propose to develop an agent-based model of a population's evacuation decision process. Simulated citizens in the model will acquire information about the approaching hurricane and 1 These are essentially linear statistical (regression) models where the dependent variable is the logistic function of the probability that a household decides to evacuate in a given time interval; independent variables typically relate to the households, the physical environment, and public information/orders. 1 decide whether and when to evacuate based on their perceptions of safety. This is not an attempt to build a simulation that can accurately forecast the collective response, but rather one that can be used to explore the dynamics of the relevant social systems. We will be performing a computational experiment to study the collective behavioral response to an extreme weather event. This is an exciting opportunity to add quantitative simulation to ISSE’s expertise in conducting research that integrates human-environment interactions with atmospheric and Earth system dynamics, and will position NCAR well to compete for upcoming NSF multidisciplinary hurricane research programs. THE BASICS OF AGENT-BASED MODELING Our approach to this problem uses an agent-based model, or ABM. In most computer simulations, such as GCMs, a physical system is represented by variables in a set of equations that describe the evolution of state over time. Continuous fields are approximated by an array of variables representing the values of the field at discrete points in space. In contrast, an ABM represents the components of the simulated system directly, as software components with behaviors mimicking the behavior of the real element. These components are placed in a virtual environment and allowed to interact with one another. Equation-based modeling works well for systems of continuous fields, while agent-based modeling is well-suited for systems composed of discrete interacting elements. For example, to model traffic with equations, we construct a mesh representing the road network, define a traffic density value at each node, and then iterate equations of motion to update the traffic density at each time step. To model it with agents, we instead create software that represents individual vehicles. Each vehicle has attributes like destination and speed, and behaviors that determine how those attributes are updated: at each time step the vehicle updates its position based on its speed, adjusts its speed based on proximity to other vehicles, preferred speed, speed limit, and other factors, and so on. We simulate traffic by placing a population of vehicles on a virtual road and allowing them to interact according to their behavioral rules. Traffic density is an emergent property of the system in this representation. Although human decision-making is too complicated to model generally, infrastructural and temporal constraints that limit the range of relevant decisions make modeling it in a limited context tractable. TRANSIMS2, for instance, clearly demonstrates that environmental (e.g., road network) and other constraints limit the extent and detail of human behavior to a degree that allows simulation. In addition to vehicular traffic, agent-based modeling has also been used successfully to study building evacuation, foot traffic in theme parks and supermarkets, NASDAQ regulations, ISP markets, operational risk factors, and other constrained social decision systems. We are not attempting to model human thought in our model; rather, we are giving our agents a simple decision-making process that approximates crisis-constrained human behavior and studying its interactions with the system of information flow in which it is embedded. Our goal is not to accurately forecast the details of an evacuation, but to gain understanding of the structural constraints imposed by system dynamics so that researchers and decision makers can develop intuition about the consequences of different evacuation policies and methods of information dissemination. AN AGENT-BASED MODEL OF HURRICANE EVACUATION In our proposed model, the agents are individual citizens who must decide whether and when to evacuate in the face of an approaching hurricane. They make this decision based on information received from the environment. Agents give different weight to information based on the communication channel through which it was received: an agent might pay more attention to evacuation orders than to weather reports, for example. In addition, the agents are linked together by a social network (their coworkers, neighbors, and friends) that provides them with awareness of decisions made by other agents. These inputs all feed into a decision-making algorithm that compares the agent's perceived risk, costs of evacuation, and time required to evacuate to determine whether to evacuate and if so, when to do so. 2 TRANSIMS, from Los Alamos National Labs, is a traffic simulation system capable of simulating the second-by-second movements of every person and vehicle through the transportation network of a large metropolitan area. See http://transims.tsasa.lanl.gov/ and http://www.transims.net for details on the TRANSIMS project. A simple demonstration of some of the principles involved is available at http://public.lanl.gov/bwb/Java/lanl/bwbush/transims/simulate.html. 2 Demographic Environment Historically Revealed Preferences Stated Preference Surveys Demographic & Socio-Economic Data Physical Environment Infrastructure Constraints Synthetic Population Social Network Crisis Scenario Cognitive Decision Model Peer Communication Infrastructure Dynamics Crisis-Behavior Decisions Public Information Information Environment Figure 1: Architectural diagram of our modeling approach Figure 2: Agent-based evacuation model structure Outcomes Performance Measures The agents have intrinsic attributes that affect their ability and inclination to evacuate, as well as their perception of various danger signals. Agents within a subpopulation share similar attribute profiles; these synthetic demographics result in different groups responding differently to the same signals. For example, one group of agents might have low evacuation costs, high weighting on evacuation orders, and loose social connectivity, similar to affluent suburbanites who are likely to evacuate early. Another group might have high evacuation costs and low receptivity to danger signals due to language barriers, but high social connectivity, much like some immigrant communities. This subpopulation's evacuation behavior will differ considerably from the first group's and will be strongly dependent on the flow of information through the social network. The action of the behavioral rules on the internal agent states and attributes can be recast as a set of iterated differential equations. An example of this transformation for our highly-simplified model can be found here: http://www.isse.ucar.edu/mcginnis/abm/proposal/simple-model-equations.pdf PRELIMINARY WORK In conceptualizing this project, we have done preliminary work to lay the foundation for its success. We have so far reviewed database entries for over 1500 articles and reports, scrutinized hundreds of abstracts, and studied hurricane impact and assessment reports that evaluate the effectiveness of evacuations for major U.S. hurricanes in the last 17 years. Our literature review has also provided us with a preliminary set of variables that we believe influence the evacuation decision and given us a sense of the data available. We have also implemented a simplified version of the model described above that demonstrates that the basic dynamic of this model produces results that are sensible and qualitatively similar to the evacuation curves presented in the post-storm assessments; see Figure 3, below. As part of the Memorandum of Understanding between UCAR and Los Alamos National Laboratory, we have helped to develop a coordinated proposal entitled "Unified Methodology for Simulating Crisis 3 Behavior"3 for LANL's Exploratory Research program. There is strong synergy between this Opportunity Fund proposal, which emphasizes quantitative application of social, cognitive, and behavioral theory in agent-based simulation, and the LANL proposal, which emphasizes a broader and more computationally intensive simulation paradigm Figure 1: Observed evacuation curve for Hurricane Opal (l) vs. simulated evacuation curve generated by simplified ABM (r). PROPOSED WORK We propose to perform the following work in this project: We will complete our review of existing research to determine the important demographic and other factors to include in communications-processing and decision-making algorithms. We will pursue collaborations with members of the cognitive psychology community like Kathleen Tierney through our local expert, Mike Page. Our review of communications factors will include existing communications models presented at the Interdepartmental Hurricane Conference that may be incorporated wholesale into the model. We will extend and elaborate the existing prototype by writing code to provide support for: a population of agents composed of subpopulations with different characteristic profiles -- that is, class, ethnicity, and other significant demographic subgroups self-perception of hazard as an input to the decision-making process economic and social constraints on the evacuation decision differential effectiveness of communication channels We will compare data produced by the completed model to data on real hurricane evacuations to evaluate its performance. We will also perform sensitivity analysis using variance-based methods (Latin hypercube and combined orthogonal array sampling) to gain insight into the dynamics of the evacuation system. RESEARCH OUTCOMES AND DELIVERABLES Social behavior has not yet quantitatively been accounted for in hurricane evacuation modeling, an oversight that contributes to loss of life. This research will begin to correct that oversight and bridge the gap between physical and social sciences, providing a tool that translates weather and climate science results like hurricane track predictions into a decision support product with meaning and value to policy makers, public planners, and laypeople confronted with the issue of hurricane evacuation. This project will generate the following knowledge products: 3 See http://www.rap.ucar.edu/~bwb/limited/20070357ER-final20060222.pdf for a copy of the proposal. 4 An agent-based simulation model of evacuation behavior that interfaces social systems with climate/weather systems, paving the way methodologically for future modeling efforts to integrate human behavior with the earth-sun-climate system. This project can act as the vanguard for one aspect of the Integrated Hurricane Program envisioned by Greg Holland, Jeff Lazo, and others. A sensitivity analysis identifying the importance of various demographic, informational, infrastructural, and policy variables that affect hurricane evacuation response. A catalog of recommendations for the behavioral survey questions most relevant to reducing uncertainties in evacuation behavior models. A computer model that is also usable in a facilitated workshop as an educational tool to aid in the development of decision-makers' intuition regarding mass evacuation behavior. BROADER IMPACTS OF THIS RESEARCH This research integrates research and education by exploring the impact of weather science results on public decisions. In addition, our model will be simple enough to be suitable for use by educators and informed laypeople. We will broaden the participation of under-represented groups in UCAR research on the impacts side by explicitly accounting for socio-economic and demographic diversity in studying the effects of hurricanes on populations. This work enhances UCAR's infrastructure for research and education by building a strong modeling collaboration with LANL, and by building the capacity for ISSE to engage in quantitative simulation of social systems and the broad modeling of weather- and climate-related human activities. It will connect with the NCAR-wide Integrated Hurricane Project, contributing social modeling capabilities to that effort. This project will require no GAUs. Results of this work will be disseminated through workshops and public availability of the model for download from the Web. This project benefits society at large by enhancing our ability to stage successful hurricane evacuations. This project will provide policy-makers with a tool that elucidates the impact that policy choices have on the populace. Products of this research will help policy-makers to better understand how their unique communities can mitigate potential human impacts of hurricanes by providing nuanced models of evacuation scenarios. In particular, results of this research will aid: Transportation-planning-oriented evacuation models, which benefit from more detailed demand models; eventually, behavioral simulations can be linked directly to traffic models The design of surveys that elicit information about critically influential behaviors Urban planning initiatives, by allowing them to better account for behavioral dynamics in evacuation. Public information dissemination, particularly to demographic groups at high risk. Our collaborators and university interactions include: Brian Bush, Los Alamos National Lab, will contribute agent-based modeling, simulation, and statistics expertise; Brian Muller, CU-Denver School of Planning, will contribute analysis of urban design factors; Sandy Johnson, LSU School of Public Health, will contribute demographic vulnerability information. Kathleen Tierney, National Hazard Center at CUBoulder, has also expressed interest in this project. REFERENCES R. Axelrod and L. Tesfatsion, "A Guide for Newcomers to Agent-Based Modeling in the Social Sciences", in L. Tesfatsion and K.L. Judd (Eds.), Handbook of Computational Economics, Vol. 2: Agent-Based Computational Economics, Elsevier, 2006; http://www.econ.iastate.edu/tesfatsi/abmread.htm E. Bonabeau, “Agent-Based Modeling: Methods and Techniques for Simulating Human Systems,” PNAS 99, Supplement. 3: 7280-7287, 2002; http://www.pnas.org/cgi/content/full/99/suppl_3/7280 E. Chin-Ping Chang, “Traffic Simulation for Effective Emergency Evacuation”, Oak Ridge National Laboratory, 2003; http://www.ictpaweb.org/publication/NCA/conferean03/Proceeding/074.pdf H. Fu and C.G. Wilmot, "Survival Analysis Based Dynamic Travel Demand Models for Hurricane Evacuation," Proc. Transportation Rsch. Board 85th Ann’l Mtg., Washington, D.C., 20-26 January 2006. C.M. Macal and M.J. North, “Tutorial on Agent-Based Modeling and Simulation”, Proceedings of the 2005 Winter Simulation Conference; http://www.informs-sim.org/wsc05papers/002.pdf 5