Speaker
Description
Advances in data collection and simulation-based inference now enable fitting complex individual-based models (IBMs) to rich, fine-grained observations of interacting systems. However, mere data fitting is not understanding: these methods do not clarify which model quantities are truly constrained by the data and which are fundamentally underdetermined. This is the problem of identifiability.
Identifiability analysis for complex, heterogeneous IBMs remains a fundamental open problem. Heterogeneous IBMs are particularly challenging due to intractable likelihoods and emergent, context-dependent behaviour. Our recent Invariant Image Reparameterisation (IIR) method, and the related Profile-Wise Analysis (PWA), provide a theoretical foundation for tackling identifiability in heterogeneous IBMs, but require an auxiliary mapping linking data features to mechanistic parameters.
Here, we present work on constructing such auxiliary mappings for IBMs and investigate what IIR can tell us about identifiability in these models.
A challenging real-world example of collective behaviour motivates our work: multi-predator feeding aggregations (MPFAs) in the Hauraki Gulf, New Zealand. We show the kinds of drone-collected data available for this application and present preliminary steps for constructing auxiliary maps from such data.
