Abstract

Generalisation of dynamical processes across natural systems is difficult because of their complexity and unobserved variables. The hope is that generalisation may be achieved if we model adequately the complexity of systems, and observe them in sufficient detail. Yet, there is still limited support for this claim. We investigate this by looking at the consistency of ecological interactions across three replicates of a three-species prey-predator system, well-observed in an artificial environment, using neural ordinary differential equations. We find that dominant interactions are consistent across the replicates, while weaker interactions are not, leading to different dynamical patterns across replicated systems. Our study hence suggests that generalisation of dynamical processes across systems may not be possible, even in simpler systems in ideal monitoring conditions. This is a problem because if we are not able to make generalisations in a simple artificial system, how can we make generalisation in the real world?