Mining stochastic cellular automata to solve density classification task in two dimensions

Density Classification Task (DCT) is a well-known problem that researchers have been tackling for more than two decades, where the main goal is to build a cellular automaton whose local rule gives rise to emergent global coordination. We describe the methods used to identify new cellular automata that solve this problem. The design of our cellular automata was carried out by a parallel genetic algorithm, specifically instantiated for this task. Our approach identifies both the neighborhood and its stochastic rule using a dataset of initial configurations that covers in a predefined and balanced way the full range of densities in DCT. We compare our results with some models currently available in the field. In some cases, our models show better performance than the best solution reported in the literature, with efficacy of 0.842 for datasets with uniform distribution around the critical density. The best-known cellular automaton achieves 0.832 in the same datasets. Tests are carried out in datasets of diverse lattice sizes and sampling conditions; we focused the analysis on the performance of our model around critical densities. Finally, by a statistical non-parametric test, we demonstrate that there are no significant differences between our identified cellular automata and the best-known model.