Practical and conceptual issues in the use of agent-based modelling for disaster management

Application of agent-based modelling technology (ABM) to disaster management has to date been limited in nature. Existing research has concentrated on extending the model structures and agent architectures of complex algorithms to test robustness and extensibility of this simulation approach. Less attention has been brought to bear on testing the current state-of-the-art in ABM for modelling real-life systems. This thesis aims to take first steps in remedying this gap. It focuses on identifying the practical and conceptual issues which preclude wider utilisation of ABM in disaster management. It identifies that insufficient attention is put on incorporating real-life information and domain knowledge into model definitions. This research first proposes a methodology by which some of these issues may be overcome, and consequently tests and evaluates it through implementation of InSiM (Incident Simulation Model), which depicts reaction of pedestrians to a CBRN (chemical, biological, radiological or nuclear) explosion in a city centre. A number of steps are conducted to obtain real-life information related to human response to CBRN incidents. This information is then used for design and parameterisation of InSiM which is implemented in three configurations. In order to identify the effects use of real-life data have on the simulation results each configuration incorporates the information at different level of complexity. The effects are assessed by comparison of the generated dispersion patterns of agents along the city centre. However, use of conventional statistical goodness-of-fit tests for assessing the degree of the difference was challenged by inhomogeneous nature of the data. Hence, alternative approaches are also adopted so that results can be qualitatively assessed. Nevertheless, the evaluation reveals significant differences at global and local level. This research highlights that incorporation of real-life information and domain knowledge into ABM is not without problems. Each time a problem was addressed, additional issues began to emerge. Most of these challenges were related to generalisation of the complex real-life systems that the model represents. Therefore, further investigations are needed at every methodological step before ABM can fully realise its potential to support disaster management