Evolvable competency and noise robustness in neural cellular automata

Shreesha and Levin (2023) gave each competent embryo an integer competency level at the start of each developmental cycle. The integer set how many successive bubble-sort swaps the embryo could perform¹. Competency levels were usually far below the total swaps needed for maximum fitness, so the sort was called restricted¹. They also varied the starting share of competent embryos in the population. When that share was 2.5%, even a competency level of 95 only reached equal percentages with hardwired embryos. At 10%, a level of 75 sufficed, and at 30% a level of 10 sufficed². Where competent embryos dominated, they did so within two or three generations².

Hartl, Risi and Levin (2024) evolved neural cellular automata (NCAs) to solve an 8×8 Czech flag problem in 25 developmental steps. They evolved them at zero noise, without and with evolvable competency³. They then raised the noise level to as much as 0.5 over lifetimes of 100 steps. Each noise level was scored as the average of 100 independent simulations, without further evolution³. Other NCAs, evolved at a competency of 0.5 and noise of 0.25 or 0.5, were exposed to noise from 0 to 1. Their competency level was also varied³. Only the correctness part of the fitness score was considered³.

Pio-Lopez, Hartl and Levin (2025) call the noise level used to corrupt the NCA update step a critical parameter. They argue that if the selective noise is too low, evolution in their case most likely finds a direct encoding of the target pattern. Cellular competencies are then largely neglected⁵. They add that too much noise prevents a good solution and limits what the cell states can specify⁵. In their aging framing, noisy updates stand for an accumulation of misdifferentiated cells⁶. The same work reports that this speeds aging but is not its cause per se⁵.

Hartl and colleagues also took an NCA evolved under one noise condition and ran CMA-ES on the same Czech flag task at higher noise. Only one generation, or at most a handful, was needed at noise levels of 0.25 and 0.5⁴. A related line of work by Etcheverry and colleagues (2025) treats gene regulatory networks as agents. They build automated tools to map the robust goal states those networks reach despite perturbations⁷.