Skip to main content
Back to timeline
arXivSource publication:

Three agents form single-file flocks from prediction alone: no alignment rule, and a temporary link cut anoints the leader

Synopsis

Three agents using Bayesian and inverse Bayesian (BIB) inference predict one another's positions and move toward those predictions, repeatedly forming leader–follower flocks and then collapsing them without any alignment rule or designated leader. Before a flock forms, one agent nearly stops incorporating observations of another, so its predicted position detaches from reality as a phantom target and carries it into the lead. Of the 1,614 flocks formed in 100 trials, the 1,312 eligible for evaluation had the three leadership criteria—front position, heading stability and flock initiation—agree on the same agent in 89% of cases, and different agents led different flocks.

Source-provided article image: Emergence and Collapse of Leader-Follower Flocks through Phantom Targets: Surprise Nullification in Bayesian and Inverse Bayesian Inference
arXiv

Interpretation

Three agents each use BIB inference to predict the positions of the other two and move toward those predicted positions, so a single-file flock with one agent in front and two following repeatedly appears and then collapses, with no alignment rule and no leader designated in advance. Conventional flocking models rely on explicit alignment rules, many adding repulsion and attraction; here the only movement rule is attraction toward the predicted positions of the other agents. Across 100 trials, BIB formed 1,614 flocks; as controls, fixed-β BIB with β₀ = 0.5, random-R BIB with randomized likelihood variance and fixed-R BIB updating only belief variance formed no flocks, while Forgetting Bayes without resets produced smoothly curving trajectories with the three agents translating side by side.

Before a flock forms, one agent's observations of another agent barely enter its update, so its predicted position for that partner is extrapolated at constant velocity away from the true position and becomes a phantom target; the agent heads straight for it and the other two trail behind. This link cut arises from two consecutive resets: one reset draws a very small likelihood variance (median 3.0×10⁻⁵), and the next draws an ordinary one, raising the likelihood-to-belief variance ratio R/P to about 2.2×10⁴ and dropping the learning rate to nearly zero. On the leaders' 1,202 cut links, the leader's turning angle fell from a median of about 105° to about 4° within 20 steps after the cut, group alignment A rose from about −0.2 to about 0.85 after 10 steps, and the leader's belief error about the cut partner reached a median of about 110 after 150 steps.

A cut persists because surprise nullification expands belief and likelihood variances by the same factor, leaving their ratio unchanged and the learning rate extremely low; when the phantom target swings back and the widened variances become excessive relative to the prediction error, a reset drives R/P far below 1, the learning rate returns to nearly 1, and the flock collapses. Without nullification, cuts did not recover during the analysis period (zero-β BIB); when only belief variance expanded, cuts recovered within a few steps (fixed-R BIB, median 2 steps and at most 5), and no flocks formed. In BIB, cut length had a median of 16 steps and 31% lasted at least 100 steps; across the 1,151 leader cuts that recovered within 2,000 steps, belief variance expanded by a median factor of 2.8×10⁴ by recovery, and 99% of cuts recovered through the only reset during the cut.

Perspective

The results come from a minimal model of three agents moving at constant speed on an unbounded plane, always observing one another and without inertia; two-dimensional position estimation uses an adaptive estimator built on the one-dimensional BIB update rule, and results are limited to the reference likelihood variance R₀ = 1. The authors propose testing whether the same leader-formation mechanism holds in larger groups, under limited fields of view, and in environments mixing state changes with outliers.

Drawing the likelihood variance at a reset from a uniform distribution is a modeling assumption rather than a derivation from the observation-noise distribution; when that distribution was bounded below at one-thousandth of the reference likelihood variance, no flocks formed. The reset criterion evaluates the current observation with the updated predictive distribution, which favors very small likelihood variances; with the pre-update predictive distribution, leader–follower flocks still formed (1,424 flocks, with the three criteria agreeing in 79% of eligible flocks) but cuts lasted longer. Observation noise is Gaussian, with neither a separate outlier-generating process nor nonstationary noise, so an adaptive advantage of this deferred attribution when state changes and outliers are mixed is not directly demonstrated.