PhD thesis
I conducted my PhD within the UMMISCO (IRD) under the supervision of Alexis Drogoul. I spent most of my time in the MSI research team in Hanoi (Vietnam). The objective of the PhD was to propose a modelling method to investigate not well defined hypotheses about complex natural systems. To do so I rooted my work within an actual application on avian influenza spread in Vietnam for which I collaborated with epidemiologists, veterinarians and official in Vietnam.
I defended my phd on the 30/09/11 in front of the following jury: Jean-François Perrot (president), Eric Ramat, Ho Thuong Vinh, Nicolas Marilleau, Alexis Drogoul (supervisor) and David Hill & François Roger (reviewers)
Abstract
Since a few years, in many different domains (e.g. sociology, ecology or economy), there is a growing trend in designing models that can be used for exploratory purposes rather than purely predictive ones. Loosely following a «complex systems» paradigm, exploratory models are often based on explicit and detailed representations of the components of the systems studied, and offer a large degree of freedom in the parameter or structural adjustments available to their final users (researchers, decision-makers or stakeholders).
They are intended to be used as a support of «as-if experiments» as they allow, through these adjustments, for the formulation of detailed hypotheses at various levels of description of the system. These hypotheses then lead to the generation of scenarios whose outcomes are explored and compared by way of repeated simulations.
Epidemiology is an interesting example of this situation. Its long modelling history can be characterised as a search for simple predictive models, but recent examples like the outbreaks of avian influenza in South-East Asia have shown their limits: without properly taking into account the interplays between social, ecological and biological dynamics to understand how this pandemic evolves, these models become useless as far as prediction goes. And when some of these dynamics are tentatively taken into account, the resulting models often become dependent on incomplete or qualitative data (for instance, the decision-making processes of social actors or the behaviours of birds), which prevents them to be used for any serious predicting purposes. As a consequence, there has been a recent shift of focus of the community on the design of exploratory models, which are meant to allow understanding the links between these dynamics, generating and studying various hypotheses, and measuring, with respect to these hypotheses, the impact of local or global policies in complex scenarios.
However, designing and using such models gives rise to serious methodological issues and existing modelling and simulation methods do not cope very well with this new type of models. When they are adapted to take their peculiarities into account, this often results in ad hoc solutions, which can barely be reused for other models in the same domain, let alone in different domains.
The objective of this thesis is to propose a domain-independent method (KIMONO) to facilitate the design and use of such exploratory models. Based on a series of examples from various domains (road traffic, social segregation, soil dynamics and more extensively epidemiology), I proceed from an account of the design requirements (taking conflicting and evolving hypotheses into account during the modelling process, producing highly-modular models, enabling an iterative modelling cycle, allowing for the collaboration of different experts and the combination of different formalisms, etc.) to a concrete proposal involving dedicated computer tools and a common accessible formalism, both aimed at facilitating the collaboration, communication and the implementation of «world models» (the name given in this proposal to open exploratory models).
The method I propose focuses on two elements: the implication of the experts and a detailed representation of the system.
Experts are at the centre of the modelling process, which starts with extensive descriptions of their knowledge, possibly reusing their formalism, and further proceeds through iterative amendments (of increasing or decreasing complexity) that they are able to evaluate and validate in interaction with the modellers. The iterative process comes to an end when the experts estimate that they have a sufficient insight of the system or when further investigations require field experiments.
Regarding the kind of representation suitable for supporting this process, I propose an adaptable and modular combination of two implementation systems: Agent-Based Modelling (ABM) and Geographical
Information Systems (GIS). I show that this combination provides for an arbitrary level of description of the components of a system, that it allows both qualitative and quantitative knowledge to be equally represented and that it supports a high level of evolution of the hypotheses during the modelling process. The interactions between modellers and experts are based on two abstractions of these implementation details, using both the ODD (Overview, Design concepts, Details) protocol for communication purposes, and the GAML modelling language for the collaborative programming of the model.
The method proposed has been applied and validated in the context of a large study undertaken in South-East Asia (especially North-Vietnam) by epidemiologists and veterinarians to understand the role of various hypotheses in explaining the recurring outbreaks of the avian influenza epidemics among domestic poultry. During a four years long interdisciplinary collaboration, several «world models» have been co- designed, implemented on the GAMA platform and used as «virtual laboratories» by experts. This collaboration, and its unique outcomes, have allowed them to test a broad range of hypotheses (especially on the local conditions of persistence), better understand the role of various spatial, ecological or social factors in the survival and propagation of the virus and reorient some of their field studies in consequence.