Constrained Neuroevolution Simulation Data

This dataset of simulation results accompanies the paper titled "Structural Convergence and Behavioral Specialization in Biologically Constrained Neural Circuits Evolved to Replicate Mouse Navigation" with the abstract as follows: How behavioral individuality is encoded in shared circuit architectures is a fundamental open question in neuroscience. We evolve 14-neuron recurrent networks (Dale's Law, sparse connectivity, quantized weights) to replicate the navigation behavior of 9 individual mice. Across 54 independent runs, all networks converge to a shared structural profile: none of 18 circuit features differ across mice, a floor imposed by the architectural constraints themselves, not by evolutionary selection (randomly initialised constrained agents return an identical 0/18 significant features). Yet circuits are behaviorally specialized: cross-mouse fitness error is 33.4% higher than own-mouse error (specialization ratio = 0.666). A permutation ablation scrambling connection topology while preserving all structural features disrupts motor output 3.53x, targeting the training mouse specifically (Wilcoxon p = 0.0039). Synaptic magnitude is uninformative and no single neuron is responsible. Behavioral identity is encoded holistically in connection topology, invisible to structural and dynamical analyses. A generalist trained on all 9 mice simultaneously shows no own-mouse permutation bias, confirming that individual specificity requires individual training targets. Instructions on using it can be found along with the accompanying code repository, available at https://github.com/pranetkhetan/constrained-neuroevolution.