To control a fish school, you need an infiltrator — but physical robotic fish are fragile, constrained in movement, and expensive. A virtual fish projected on a screen is none of these things. The problem: real fish frequently ignore the virtual one.

Reinforcement learning solves this not by making the virtual fish more convincing, but by making it more strategic. The RL agent learns when to move, how fast, and in what direction to maximize the school’s displacement toward a target — accounting for the fact that the fish will often not follow. The policy succeeds not because every fish responds every time, but because the aggregate effect of occasional responses, compounded over time, steers the school.

The trained virtual fish significantly outperforms both no-stimulus controls and hand-designed heuristic movement patterns. The learned strategy exploits the school’s own social dynamics: when even a few fish respond to the stimulus, their movement influences neighbors through the school’s internal alignment rules. The virtual fish doesn’t need to convince the whole school — just enough individuals to trigger a social cascade.

This is influence without presence. The virtual fish has no physical reality, exerts no forces, and can be ignored with impunity. Its power comes entirely from timing — appearing in the right place at the right moment to tip the school’s collective decision. The shepherd doesn’t need to be real. It just needs to understand when the flock is ready to turn.