Disentangling intrinsic motion from neighbourhood effects in heterogeneous collective motion

TL;DR

This paper presents a novel method to classify heterogeneous agents in collective motion by explicitly accounting for neighborhood effects, improving the accuracy of inferring intrinsic properties from observed movement patterns.

Contribution

The paper introduces a new approach that incorporates neighborhood effects into classification, enabling better inference of individual intrinsic properties in heterogeneous collectives.

Findings

Significant improvement in classification accuracy using the proposed method.

Explicitly modeling neighborhood effects helps distinguish agent types with identical observed velocities.

The approach effectively infers heterogeneity levels in complex collective systems.

Abstract

Most real world collectives, including active particles, living cells, and grains, are heterogeneous, where individuals with differing properties interact. The differences among individuals in their intrinsic properties have emergent effects at the group level. It is often of interest to infer how the intrinsic properties differ among the individuals, based on their observed movement patterns. However, the true individual properties may be masked by emergent effects in the collective. We investigate the inference problem in the context of a bidisperse collective with two types of agents, where the goal is to observe the motion of the collective and classify the agents according to their types. Since collective effects such as jamming and clustering affect individual motion, an agent's own movement does not have sufficient information to perform the classification well: a simple observer algorithm, based only on individual velocities cannot accurately estimate the level of heterogeneity of the system, and often misclassifies agents. We propose a novel approach to the classification problem, where collective effects on an agent's motion is explicitly accounted for. We use insights about the physics of collective motion to quantify the effect of the neighbourhood on an agent using a neighbourhood parameter. Such an approach can distinguish between agents of two types, even when their observed motion is identical. This approach estimates the level of heterogeneity much more accurately, and achieves significant improvements in classification. Our results demonstrate that explicitly accounting for neighbourhood effects is often necessary to correctly infer intrinsic properties of individuals.