Visual and semi-automated exploration of an archaeological model: calibrating a model of spatial reorganization in North-Western Europe, A.D. 800 to 1100
Résumé
Most often, when building a model, the modeller has to choose between two types of methods for
evaluating its model. (1) The “face validation”, based on a visual and qualitative assessment of the model outputs, allows for a quick and direct feedback of the model behaviour considering the different values of parameters used. (2) More recently, and to have a more pervasive method of exploration, many modellers leant on High-Perfomance Computing to develop automated ways of exploring models often based on full factorial designs or on more parsimonious methods like genetics algorithms.
In this presentation, we advocate for combining the efficiency of the automated explorations and the
qualitative feedback of face validation, in particular in a context of strong interdisciplinarity involving
geographers, archaeologists and historians.
Our case study is a model of settlement system evolution over the long time that aims at a better
understanding of the social and spatial processes that happened in North-Western Europe between 800 and 1100 A.D.
This period is characterized by a major change in the settlement system, to which we refer as the “800-1100 transition” (Tannier & al., 2014). The peasant households, mostly sprawled around 800, tend to cluster around functional attractors, mainly castles and churches. The clustering and fixation of households are at the origin of villages and small towns. With the Gregorian Reform, which led to a stronger religious control, a grid of parishes emerged around the churches.
We have introduced these processes in an agent-based model built with the GAMA platform, with the aim to test hypotheses on triggers and mechanisms that produced the observed changes. For example, it should help understanding if population growth is essential to the observed clustering.
Due to the lack and uncertainty of empirical data, the model mostly relies on expert qualitative knowledge.
Consequently, we cannot define fitness functions or simple output indicators that would allow a fully
automated evaluation of the model. Moreover, the model contains a large number of parameters. In order to find a first set of parameters so that the model reproduces a situation that experts would validate, we used two complementary approaches.
First, we carried out a semi-automated exploration based on rough parameters values, in order to restrain the parameters’ space to analyse. To be parsimonious, even in the computation power required, we limit this exploration to the most obvious parameters affecting the simulation results.
We then used a visual approach, based on multiple display types to allow the specialists to understand and assess the output of an experiment. For this, we used GeoVisual Analytics methods introduced in a custom platform dedicated to the analysis of long-time spatio-temporal simulation data.
This communication aims at presenting the way those seemingly opposed parametrization methods have been articulated, and the simulation results thus obtained.