AI-driven automated discovery tools reveal diverse behavioral competencies of biological networks

Visualizations in Figure 9 show: (a-b) the oscillators discovered by random search and curiosity search (gradient descent did not find any oscillator in this example) in the A,ω,b space, (c) the corresponding diversity (using this time a binning-based space coverage measure with 203 bins as the space is 3-dimensional), (d) the evolution of the training loss L throughout the n=5000 trials for the three exploration strategies, (e-f-g) the corresponding best discoveries (for which L is minimal) for the three exploration strategies, and (h-i) the local training loss and resulting finetuning of the best discoveries with gradient descent.

Statistical analyses were performed using custom code in Python 3.9 (relevant libraries include jax 0.4.8 and scipy 1.10 and plotly 5.16.1). Welch’s two-sample t-tests reported in Figure 3 are two-tailed, and additional details on statistical analyses are provided in the figure legend.