Abstract
Individuals exposed to similar environments often develop consistent behavioral differences, yet how learning generates individuality remains unclear. We study agents whose neural connections develop through unsupervised Hebbian learning during pursuit of a moving target. All agents begin with identical networks without recurrent connections and neurons tuned to world-fixed directions. The target coactivates neurons with overlapping directional receptive fields while leaving more distant neurons relatively inactive. Hebbian learning converts this activity pattern into local excitation and longer-range inhibition. Stochastic differences in target motion and neural activity during learning produce individual-specific departures from each network’s average center–surround profile. We then freeze the connections and repeatedly test agents with two equally attractive targets, a task absent from training. Individual agents tend to spend more time near one target than the other, with preferences differing among agents but remaining repeatable across trials. Replacing each network with its average connectivity profile strongly reduces preference strength and leaves little repeatability. Rearranging which connections are stronger or weaker than average, while preserving the amount of variation and average profile, retains repeatable preferences and improves tracking in a separate moving-target test. However, preferences confer no general advantage when targets differ in value. Under standard sensory and inhibitory settings, repeatable preferences and accurate choices coexist only at intermediate neural noise levels tested. A matched egocentric model, with tuning relative to the agent’s heading, shows no detectable target-side repeatability. Hebbian learning can therefore generate individual spatial preferences, whose expression depends on how the network represents space.