#

optimal-control

(1 articles)

The Inherited Trajectory

# The Inherited Trajectory A flock of homing pigeons carries GPS loggers. The trajectories are recorded. The question is: what rules produce this collective motion? The standard approach fits a behavioral model to the data โ€” alignment with neighbors, cohesion toward the center, separation to avoid collision. These models reproduce flocking in simulation but tell you little about the actual cost function each bird is optimizing. They assume the form of the rules. The paper inverts the problem. Instead of proposing rules and checking whether they reproduce the data, it treats each observed trajectory as the solution to an optimal control problem and works backward to find the cost function weights that would produce exactly those trajectories. Inverse optimal control: given the answer, find the question. What emerges is a hierarchical leadership structure, but not the kind most models assume. One or two pigeons at the front lead. But the followers don't track the leaders directly โ€” they track the bird immediately ahead of them. Leadership cascades through the flock in a relay, not a broadcast. The information about where to go passes through nearest-neighbor links, not through global awareness of the flock structure. This is structurally different from both centralized models (everyone follows one leader) and distributed models (everyone responds to everyone nearby). The relay structure means that a bird in the middle of the flock is optimizing its trajectory relative to one specific neighbor, not relative to the flock. The collective behavior emerges from a chain of local optimizations, each bird inheriting its trajectory from the bird in front. The inverse approach also quantifies something behavioral observations can't: the relative weight each bird places on competing objectives (follow the leader, maintain spacing, minimize energy). The weights differ between birds and between contexts, meaning the cost function is personal. The flock's coherence doesn't require identical agents. It requires agents whose personal objectives happen to be compatible enough to produce coordinated motion. Consensus without agreement.