Multi-scale Inference of Interaction Rules in Animal Groups Using Bayesian Model Selection

Richard P. Mann, Andréa Perna, Daniel Strömbom, Roman Garnett, James E. Herbert‐Read, David J. T. Sumpter, Ashley J. W. WardPLoS 2012 · DOI · OpenAlex

This paper, published in PLoS Computational Biology, provides a rigorous framework for understanding how individual interactions lead to collective behavior in animal groups. By studying glass prawns, the authors demonstrate that the "memory" of past encounters is the missing link required to reconcile individual-level rules with group-level patterns.

TL;DR

Standard models of collective motion often fail to bridge the gap between individual behavior and group dynamics. By applying Bayesian model selection to glass prawn movements, researchers found that individuals "remember" past interactions; only these non-Markovian models can accurately predict both fine-scale individual turns and the large-scale emergence of group-wide alignment.

What Problem Does This Paper Address?

In the study of collective motion (like bird flocks or fish schools), researchers typically take one of two paths:

  1. Top-down: Create a simulation and tweak parameters until the group looks "right" (e.g., matching the polarization or density of a real flock).
  2. Bottom-up: Track individuals at high resolution to infer the mathematical rules of their interactions.

The problem is that these two scales rarely align. A model that perfectly predicts how an individual prawn reacts to its neighbor often fails to produce a cohesive group when simulated. Conversely, "Mean Field" models that assume everyone interacts with everyone else can replicate group patterns but are biologically unrealistic at the individual level. The authors sought a model that is multi-scale consistent—one that works at both the "micro" and "macro" levels.

Methods

The researchers placed glass prawns (Paratya australiensis) in an annular (ring-shaped) arena and tracked their movements. To analyze the data, they employed several sophisticated techniques:

  • Hidden Markov Models (HMM): Raw tracking data is noisy. The authors used an HMM to classify movements into a clean binary sequence of clockwise (CW) or anti-clockwise (CCW) orientations.
  • Bayesian Model Selection: They compared 11 different models ranging in complexity:
  • Null/Non-spatial: Random turning.
  • Mean Field: Individuals react to the whole group.
  • Spatial Markovian: Individuals react only to current neighbors within a specific zone.
  • Non-Markovian: Individuals have a "memory" (an auto-regressive turning intensity) that integrates past interactions over time, decaying at a specific rate ($d$).
  • Marginal Likelihood: Instead of just looking at the "best fit" parameters, they integrated over the entire parameter space to see which model structure was most probable given the data, naturally penalizing overly complex models (Occam’s Razor).

Datasets

The study used experimental data from glass prawns:

  • Subjects: Groups of 3, 6, and 12 prawns.
  • Setup: An annular arena (200 mm external diameter) filled with freshwater.
  • Tracking: Video recorded at 15 Hz, downsampled to 7.5 Hz for computational efficiency.
  • Benchmark: The primary "training" data came from the 6-prawn trials, which offered the best balance of interaction density and tracking accuracy.

Key Insights and Ideas

  • Memory is Mandatory: The most significant "aha" moment was that standard Markovian models (where the next move depends only on the current state) could not reproduce group alignment. Prawns need to "integrate" their experiences. If a prawn hits several others moving the opposite way, its "urge" to turn builds up over time rather than being an instantaneous reaction.
  • Physical Contact vs. Vision: The inferred interaction zone was very small—approximately $\pi/10$ radians, which corresponds to about half a prawn's body length. This suggests that for these prawns, collective motion is driven by physical touch and short-range sensing rather than long-range visual cues.
  • The Failure of Spatial Markov Models: Interestingly, the models that most closely resembled classic "Self-Propelled Particle" models (like the Vicsek model) were excellent at predicting individual turns but failed miserably at producing group-level order in simulations.

Key Results

  • Model Performance: The non-Markovian models (D1–D4) significantly outperformed all others in marginal likelihood.
  • Global Consistency: Model D3 (a non-Markovian model with a symmetric interaction zone) provided the best fit. It achieved a low Kullback-Leibler (KL) divergence of 1.46 bits when comparing simulated group distributions to experimental ones.
  • Phase Transition: The data confirmed a "phase transition" where increasing prawn density leads to a sudden onset of collective motion in one direction. The non-Markovian models were the only ones to accurately replicate the timing and "sharpness" of this transition.

Limitations and Future Work

  • Dimensionality: The study was conducted in a 1D-like annular arena. While this simplifies the math, it remains to be seen how "memory" effects scale in 2D or 3D environments where animals have more freedom to avoid collisions.
  • Binary Simplification: The movement was reduced to a binary CW/CCW choice. Future work could incorporate more nuanced behaviors like speed changes or pauses.
  • Weighting Global Consistency: The authors noted that while Bayesian selection identifies the best individual-level model, a more robust future approach might formally incorporate a "weighting" for how well that model performs at the group scale during the selection process.

Summary updated

Abstract

Inference of interaction rules of animals moving in groups usually relies on an analysis of large scale system behaviour. Models are tuned through repeated simulation until they match the observed behaviour. More recent work has used the fine scale motions of animals to validate and fit the rules of interaction of animals in groups. Here, we use a Bayesian methodology to compare a variety of models to the collective motion of glass prawns (Paratya australiensis). We show that these exhibit a stereotypical 'phase transition', whereby an increase in density leads to the onset of collective motion in one direction. We fit models to this data, which range from: a mean-field model where all prawns interact globally; to a spatial Markovian model where prawns are self-propelled particles influenced only by the current positions and directions of their neighbours; up to non-Markovian models where prawns have 'memory' of previous interactions, integrating their experiences over time when deciding to change behaviour. We show that the mean-field model fits the large scale behaviour of the system, but does not capture fine scale rules of interaction, which are primarily mediated by physical contact. Conversely, the Markovian self-propelled particle model captures the fine scale rules of interaction but fails to reproduce global dynamics. The most sophisticated model, the non-Markovian model, provides a good match to the data at both the fine scale and in terms of reproducing global dynamics. We conclude that prawns' movements are influenced by not just the current direction of nearby conspecifics, but also those encountered in the recent past. Given the simplicity of prawns as a study system our research suggests that self-propelled particle models of collective motion should, if they are to be realistic at multiple biological scales, include memory of previous interactions and other non-Markovian effects.
BibTeX
@inproceedings{mann2012multi,
  author = {Richard P. Mann and Andréa Perna and Daniel Strömbom and Roman Garnett and James E. Herbert‐Read and David J. T. Sumpter and Ashley J. W. Ward},
  title = {{Multi-scale Inference of Interaction Rules in Animal Groups Using Bayesian Model Selection}},
  booktitle = {PLoS},
  year = {2012},
  doi = {10.1371/journal.pcbi.1002308},
}

Loading related content...

Figure