Evolving Many Worlds: Towards Open-Ended Discovery in Petri Dish NCA via Population-Based Training

A team of researchers including Uljad Berdica, Jakob Foerster, Frank Hutter, and Arber Zela has published a paper introducing PBT-NCA, a novel meta-evolutionary algorithm designed to tackle a core challenge in artificial life: generating sustained, open-ended complexity. The system works on Petri Dish Neural Cellular Automata (PD-NCA), which are differentiable multi-agent systems known for rich self-organization but also for being highly sensitive to hyperparameters, often collapsing into frozen patterns or noise.

PBT-NCA evolves a population of these PD-NCAs using a composite objective function that rewards both historical behavioral novelty and contemporary visual diversity. This continuous evolutionary pressure actively penalizes monocultures and dead states, forcing the system to operate persistently at the 'edge of chaos'—a state of effective complexity that is neither globally ordered nor random. The result is the spontaneous, long-term emergence of a plethora of lifelike phenomena that were not explicitly programmed.

Strikingly, the digital substrate autonomously discovers diverse and sophisticated morphological survival strategies. The researchers observed highly regular, coordinated periodic waves; spore-like scattering where homogeneous groups eject cell-like clusters to colonize distant territories; and fluid, shape-shifting macro-structures that migrate across the petri dish while maintaining stable boundaries. These behaviors emerge purely from local spatial competition driven by the algorithm's reward for novelty and diversity, marking a significant step toward true open-ended discovery in simulated environments.