Managing structural variability in agent-based models with feature coherence graphs

Abstract

Exploratory modelling requires not only parametric but also structural and representational variability. The ability to systematically vary and experiment with multiple model features is paramount for problems that involve uncertainty and representational ambiguity. To facilitate such variation, a coherence-driven strategy is introduced for the management and evaluation of feature selections. The feature coherence specification and its associated metaprogramming system open new avenues for programmable abductive model building, as well as model introspection while automating the selection of model variants in the context of an evolving analysis. As learning takes place through simulation experiments, an increasingly accurate feature coherence model emerges to serve as an explanatory model that reveals which features cohere and are conducive to generating targeted systemic behaviours.