SWIFT: Simulations With Intelligence for Fire Training Carole Adam1,2 , Elise Beck3,2 and Julie Dugdale1,4,5 {carole.adam,julie.dugdale}@imag.fr [email protected] 1 Grenoble Informatics Laboratory, France 2 University Joseph Fourier, Grenoble, France 3 PACTE laboratory, Grenoble, France 4 University Pierre Mendes-France, Grenoble, France 5 University of Agder, Norway

1

Statement of topic

Track 2:

Analytical Modeling and Simulation

The state of Victoria in Australia faces bushfires of varying seriousness every summer. On the “Black Saturday”, 7th of February 2009, particularly violent bushfires resulted in 173 victims and burnt acres of bush. Despite regular awareness campaigns, the state policy titled “prepare, stay and defend, or leave early” was not followed by the population. Reports from the Royal Commission [Teague et al., 2009] found that most people would still wait and see. While the emergency managers only broadcast general alert messages, the population would wait for a personal evacuation order, or the immediate proximity of the fire to trigger their escape. Therefore most victims were people evacuating at the last minute. Researchers tried to explain this unexpected ”wait and see” behaviour and found that information overload would lead to ”paralysing indecisiveness” [McNeill et al., 2014]. These results suggest that more targeted information and alert messages might help in reducing the casualties, but it is hard to predict their actual impact ahead of time. Computer modeling and simulation offers a powerful tool to evaluate the efficiency of such emergency management policies and tailor them without waiting for an actual bushfire to happen, so without any human life at stake, and with a great degree of control over all contextual variables, allowing tus o reproduce the exact same settings as many times as needed, or to experiment with different scenarios. Compared to mathematical models, multi-agent models offer a higher level description of human behaviour, more faithful to reality. They allow to study social systems at various levels of abstraction depending on the complexity of the agents used. Multi-agent modeling and simulation is already used for decision support in crisis management. However the lack of real geographical data limits the validity of their results; to be of real use such simulations should be spatialised, relying on real Geographical Information Systems (GIS) data. Another limit of existing simulations is the over-simplification of agents behaviour. On the contrary in crisis management and response, we need realistic cognitive agents, not only reacting to environmental stimuli, but also capable of complex decision-making in an uncertain world, influenced by their emotions, and interacting with other individuals and groups (family, community). To summarise, the SWIFT project is relevant to the ISCRAM conference themes by providing an agent-based model and simulation of the population’s response to bushfires: people often behave in unpredictable and irrational ways in such crisis due to emotions and stress, and emergency managers can use this model to better understand the underlying decision making processes. Furthermore, the realism of our model, ground on actual geographical data and field surveys, will ensure the validity and significance of its results in order to improve the management and response to bushfires in the future.

2

Abstract

In this context, the SWIFT project aims at modeling individual people’s decision-making and behaviour when facing a bushfire. We have previously developed the AMEL model of pedestrian mobility after an earthquake [Truong et al., 2013, Beck et al., 2014], and successfully applied it to two different contexts: a suburb of Beyrouth in Lebanon, and the town of Mendoza in Argentina. Our work follows a five-step methodology (cf Figure 1) with many iterations between all phases to ensure the validity of the model and its fit to the data. We use the same methodology for SWIFT.

Figure 1: Methodology

We are now gathering data specific to the context of the Victorian bushfires, in different ways: Studying the existing reports and surveys conduced after the 2009 bushfires; designing a questionnaire for surveying the population in bushfire-exposed areas, focused on questions not already answered in the literature (in particular we are interested in the use of social media as an information source during a bushfire, and in the influence of emotions on behaviour); and interviewing emergency managers about their policies. The data we gathered so far shows the importance of human factors: emotional decisions under stress, inter-individual communication and imitation, and information-seeking behaviours. The SWIFT project therefore requires complex cognitive agents, endowed with mental attitudes (Beliefs, Desires, Intentions BDI) determining their behaviour. However the implementation of AMEL was done in GAMA [Grignard. et al., 2013], an opensource simulation platform allowing the integration of GIS data and the easy development of agent models by non-specialists, but with relatively simplistic agent architectures. We are investigating two methods of integrating BDI agents into GAMA simulations: extending GAMA with a new plugin providing a BDI agent architecture; or using RMIT middleware for combining Agent-Based modeling platforms with BDI development frameworks [Padgham et al., 2014]. Such work will be significant for all sorts of social simulations involving individual decision-making. Once our model is developed and validated on the field data, we will run simulations on various scenarios, in consultation with emergency managers in Victoria, in order to help them improve their policies. In particular our scenarios will test: the use of different media (including social media) to broadcast information and alerts, and how to combine these with ”traditional” media; the specificity of the message content: targeting a geographical area, age groups, gender groups, individual people; and the influence of various message contents on people’s emotions, in an attempt to reduce information overload and stress while still keeping people informed. To conclude, the SWIFT project has just started but we can already present some interesting behavioural data, a methodology for integrating more intelligent agents into social simulations, and application scenarios for our simulation. In future work we plan to transform this agent-based model into a serious game that can be used by the population to practice their bushfire awareness and preparedness, and also by the emergency managers to gather data to better understand the population behaviour and tailor their information policies accordingly.

References Elise Beck, Julie Dugdale, Hong Van Truong, Carole Adam, and Ludivine Colbeau-Justin. Crisis mobility of pedestrians: from survey to modelling, lessons from lebanon and argentina. In Information Systems for Crisis Response and Management in Mediterranean Countries (ISCRAM-Med), volume 196 of Lecture Notes in Business Information Processing, pages 57–70. Springer, Toulouse, France, 15-17 October 2014. A. Grignard., P. Taillandier, B. Gaudou, N.Q. Huynh, D.-A. Vo, and A. Drogoul. Gama v. 1.6: Advancing the art of complex agent-based modeling and simulation. In PRIMA, 2013. Ilona McNeill, Patrick Dunlop, Timothy Skinner, and David Morrison. Information processing under stress: Community reactions. Technical report, Bushfire CRC, Australia, 2014. http://goo.gl/TkcC4d. Lin Padgham, Kai Nagel, Dhirendra Singh, and Qingyu Chen. Integrating BDI agents into a matsim simulation. In European Conference on Artificial Intelligence (ECAI), pages 681–686, 2014. Bernard Teague, Ronald McLeod, and Susan Pascoe. Final report. Technical report, 2009 Victorian Bushfires Royal Commission, 2009. http: //goo.gl/L5BCHf. Hong Van Truong, Elise Beck, Julie Dugdale, and Carole Adam. Developing a model of evacuation after an earthquake in Lebanon. In ISCRAM Vietnam, 2013.