year: 2023
paper: https://www.nichele.eu/ALIFE-DistributedGhost/2-Pontes.pdf
website:
code:
connections: NCA, criticality, self-organized criticality, Stefano Nichele
Short paper that demonstrates critical NCAs.
Achieving Criticality in Artificial Systems
One approach is to optimize a simple and deterministic dynamical system towards criticality. Neural Cellular Automata (NCA) are being explored for this purpose. Evolutionary algorithms, like cma-es, are used to evolve the parameters of the NCA to achieve criticality, identified by power-law distributions in avalanche sizes, durations, and areas. The goal is to achieve critical behavior, which is identified by finding power-law distributions in the measured characteristics (size, duration, area) of activity avalanches within the NCA. This paper evolves a deterministic 1D NCA with binary states and communication channels, optimizing a fitness function based on how well the avalanche distributions fit power laws.
Defining Avalanches
In the context of Cellular Automata studied for criticality, an “avalanche” refers to a connected cluster of active states that propagates through space and time during the system’s evolution. Criticality is often identified by analyzing the statistical distributions of these avalanches, specifically looking for power-law relationships in their:
Size: The total number of unique cells that become active at least once during the avalanche.
Duration: The total number of consecutive time steps the avalanche lasts, from its start until no cells belonging to it are active.
Area: The total number of active site instances (cell active at a specific time step) summed over the entire duration and spatial extent of the avalanche.istributions fit power laws.
Criticality and Computation
Systems at criticality are hypothesized to have optimal computation capabilities. Evolving artificial systems like NCAs to operate at criticality could lead to more adaptive AI systems, potentially applicable in paradigms like reservoir computing.