Article Preview
TopIntroduction
Each summer in Australia, bushfires burn many hectares of forest, causing deaths, injuries, and destroying property. Societies can manage such crisis and emergency situations in several ways: adopt urban and territory planning policies to reduce the risks (e.g. forbid construction in exposed areas); raise awareness and prepare the population in advance; or create efficient emergency management policies to deal with crises when they happen. Modelling and simulation offer tools to test the effects and complex interactions of these different strategies without waiting for an actual crisis to happen, without putting human lives at risk, with limited cost, and with a great degree of control on all conditions and the possibility of reproducing exactly the same situation as many times as needed.
When modelling human behaviour, mathematical, equation-based models are too limited (Parunak, Savit, & Riolo, 1998) whereas agent-based models offer many benefits (Bonabeau, 2002). They allow capturing emergent phenomena that characterise such complex systems; they provide an intuitive and realistic description of their behaviour; and they are flexible, offering different levels of abstraction by varying the complexity of agents. Agent-based modelling and simulation platforms over various architectures of different complexity for the agents: reflex or reactive agents are very simplistic, reacting to environ-mental stimuli without any long-term reasoning; finite-state machines require scripting all of the possible states of the agents and the corresponding transitions; cognitive agents offer a more flexible description of behaviours in terms of goals and plans.
In particular the BDI (Belief, Desire, Intention (Rao & George, 1991)) architecture is very sophisticated and realistic, grounded in the philosophy of human rationality (Bratman, 1987), and linked with emotions (Adam, Herzig, & Longin, 2009). Such realism of the human behaviour model is important for the simulation to produce valid results (van Ruijven, 2011). For these reasons and as previously discussed in (Adam & Gaudou, 2016a), the BDI architecture is therefore more adapted for crisis situations that involve complex individual decision-making, influenced by emotions (sometimes causing irrational actions), and by the social context (effect of group, family, etc.). According to (Norling, 2004), BDI also provides an adapted level of abstraction to describe human behaviour in terms of folk psychology, which is the preferred level of description for humans. It therefore addresses the problem of the scarcity of (quantitative) behavioural data by allowing the use of qualitative data such as witness statements or expert reports.
Despite these advantages that make it very suitable for social simulation, BDI has had limited use in this field due to the lack of adapted tools to harness its complexity (Adam & Gaudou, 2016a). In previous work (Adam, Danet, Thangarajah, & Dugdale, 2016; Adam, Beck, & Dugdale, 2015; P. Taillandier, Bourgais, Caillou, Adam, & Gaudou, 2016) we have described how two new tools could be used to de-velop BDI agent-based models from interviews. We illustrated how to use these tools by turning an existing model of the behaviour of the Australian population in bushfires (with a finite-state machine architecture (Adam & Gaudou, 2016b, 2017)) into a BDI model. In this paper we now want to com-pare these two models, addressing the same problem with different architectures, using both objective and subjective criteria. We believe that such model comparison is important to further justify the interest of BDI models.