PBT-NCA: Evolving Neural Cellular Worlds

• PBT-NCA is a meta-evolutionary framework that uses population-based training to continuously evolve rich, self-organized cellular ecologies.
• It overcomes hyperparameter sensitivity by inheriting entire training states and employing exploit–explore replacements, ensuring sustained novelty.
• The method combines historical behavioral novelty and contemporary visual diversity to maintain ecosystems at the edge of chaos for open-ended evolution.

Searching arXiv for the specified PBT-NCA paper and closely related PD-NCA context. PBT-NCA denotes Population-Based Training for Petri Dish Neural Cellular Automata, a meta-evolutionary framework for discovering sustained, open-ended complexity in differentiable multi-agent cellular ecologies. It was introduced to address a recurrent failure mode of standard Petri Dish Neural Cellular Automata (PD-NCA): although PD-NCA can exhibit rich self-organization driven by spatial competition, it is highly sensitive to hyperparameters and often collapses into frozen equilibria, structureless noise, or monocultures. PBT-NCA replaces single-run optimization with a population-level exploit-explore process and scores worlds by a composite objective that combines historical behavioral novelty with contemporary visual diversity, thereby maintaining continuous pressure toward nontrivial ecological and morphological differentiation (Berdica et al., 13 Apr 2026).

1. Problem formulation and conceptual scope

PBT-NCA is situated in artificial life and open-ended evolution research, where the central problem is how to generate sustained, open-ended complexity from local interactions. In the underlying PD-NCA substrate, desired behavior is not a fixed target. Instead, the system is continually adaptive and competitive, so the outer optimization process determines which ecological regime survives. Small changes in learning rate, batch size, update frequency, or temperature can therefore redirect dynamics toward qualitatively different attractors, including dead worlds, static coexistence, or high-entropy disorder (Berdica et al., 13 Apr 2026).

The framework is designed to move beyond single-target optimization and manual hyperparameter tuning. Its stated goal is not merely to optimize one world for one trajectory, but to create a search process that continues to produce novel, structurally diverse, dynamically active worlds over long horizons. In that sense, PBT-NCA is presented as a practical mechanism for nonstationary discovery rather than convergence to a single optimum.

A common misconception is to treat PBT-NCA as only a hyperparameter-search wrapper for PD-NCA. The method is broader than that. Each population member is an evolving ecosystem, and the outer loop selects not just parameter settings but lineages of worlds with inherited weights, optimizer state, world state, and hyperparameters. This makes the method meta-evolutionary in a stronger sense than standard hyperparameter optimization.

2. PD-NCA substrate and local competitive dynamics

A PD-NCA world is a differentiable, multi-agent cellular automaton-like ecosystem with world state

$X^t \in \mathbb{R}^{H \times W \times C},$

where each cell $(u,v)$ stores a feature vector

$$x_{u,v}^t = [\mathbf{a}_{u,v}^t;\mathbf{d}_{u,v}^t;\mathbf{h}_{u,v}^t].$$

The channels are partitioned into attack channels $\mathbf{a}_{u,v}^t \in \mathbb{R}^{C_a}$, defense channels $\mathbf{d}_{u,v}^t \in \mathbb{R}^{C_d}$, and hidden state $\mathbf{h}_{u,v}^t \in \mathbb{R}^{C_h}$. The system also maintains an aliveness mask

$A^t \in \mathbb{R}^{N \times H \times W},$

which tracks which of the $N$ agents occupies each cell (Berdica et al., 13 Apr 2026).

Each living agent $k$ observes its local Moore neighborhood $\mathcal{N}_{u,v}(X^t)$ and proposes an update

$(u,v)$0

In addition, an environmental agent $(u,v)$1 injects background noise or decay,

$(u,v)$2

which supplies persistent background pressure against static triviality.

Competition is mediated by pairwise cosine-like interactions,

$(u,v)$3

The competitive strength of agent $(u,v)$4 at cell $(u,v)$5 is defined as

$(u,v)$6

followed by softmax normalization,

$(u,v)$7

The resulting cell update is

$(u,v)$8

The aliveness state is then thresholded as

$(u,v)$9

The threshold $$x_{u,v}^t = [\mathbf{a}_{u,v}^t;\mathbf{d}_{u,v}^t;\mathbf{h}_{u,v}^t].$$0 allows limited coexistence and keeps boundaries dynamically active rather than enforcing winner-take-all occupancy. This substrate is the dynamical object over which PBT-NCA performs selection.

3. Population-based training as a meta-evolutionary outer loop

PBT-NCA maintains a population

$$x_{u,v}^t = [\mathbf{a}_{u,v}^t;\mathbf{d}_{u,v}^t;\mathbf{h}_{u,v}^t].$$1

where each individual $$x_{u,v}^t = [\mathbf{a}_{u,v}^t;\mathbf{d}_{u,v}^t;\mathbf{h}_{u,v}^t].$$2 is a whole world containing four inherited components: NCA parameters, optimizer state, world state, and learning hyperparameters (Berdica et al., 13 Apr 2026). This is explicitly a Lamarckian scheme: offspring inherit not only learned parameters but also training context.

The reported default meta-training configuration is as follows:

Quantity	Value
Population size $$x_{u,v}^t = [\mathbf{a}_{u,v}^t;\mathbf{d}_{u,v}^t;\mathbf{h}_{u,v}^t].$$3	30
Meta-iterations $$x_{u,v}^t = [\mathbf{a}_{u,v}^t;\mathbf{d}_{u,v}^t;\mathbf{h}_{u,v}^t].$$4	500
Inner rollout $$x_{u,v}^t = [\mathbf{a}_{u,v}^t;\mathbf{d}_{u,v}^t;\mathbf{h}_{u,v}^t].$$5	12
Exploit-explore interval $$x_{u,v}^t = [\mathbf{a}_{u,v}^t;\mathbf{d}_{u,v}^t;\mathbf{h}_{u,v}^t].$$6	5
Replacement fraction $$x_{u,v}^t = [\mathbf{a}_{u,v}^t;\mathbf{d}_{u,v}^t;\mathbf{h}_{u,v}^t].$$7	0.25

At each meta-iteration, every world is rolled out for $$x_{u,v}^t = [\mathbf{a}_{u,v}^t;\mathbf{d}_{u,v}^t;\mathbf{h}_{u,v}^t].$$8 inner steps and scored by a composite objective. The top descriptors are inserted into an archive, and every $$x_{u,v}^t = [\mathbf{a}_{u,v}^t;\mathbf{d}_{u,v}^t;\mathbf{h}_{u,v}^t].$$9 iterations the system performs exploit-explore replacement: the bottom $\mathbf{a}_{u,v}^t \in \mathbb{R}^{C_a}$0 fraction of worlds is discarded and replaced by mutated copies of elite worlds (Berdica et al., 13 Apr 2026).

The replacement operator combines copying, crossover, hyperparameter mutation, and weight perturbation. Let

$\mathbf{a}_{u,v}^t \in \mathbb{R}^{C_a}$1

For each replacement, the algorithm first deep-copies the parent world into the child, including parameters, optimizer state, world state, and hyperparameters. It then applies independent crossover to each hyperparameter with probability $\mathbf{a}_{u,v}^t \in \mathbb{R}^{C_a}$2; if crossover occurs, the copied parent value is retained with probability $\mathbf{a}_{u,v}^t \in \mathbb{R}^{C_a}$3, otherwise the child reverts to its pre-copy value. Each hyperparameter is further perturbed with probability $\mathbf{a}_{u,v}^t \in \mathbb{R}^{C_a}$4, using multiplicative factors $\mathbf{a}_{u,v}^t \in \mathbb{R}^{C_a}$5 or $\mathbf{a}_{u,v}^t \in \mathbb{R}^{C_a}$6 followed by clipping to range. Finally, Gaussian noise is added to all trainable weights.

This design turns hyperparameter sensitivity into an evolutionary search dimension. The framework does not attempt to eliminate sensitivity; instead, it exploits sensitivity to traverse ecological regimes that fixed-hyperparameter optimization fails to sustain.

4. Composite objective, archive structure, and diagnostic metrics

The score of world $\mathbf{a}_{u,v}^t \in \mathbb{R}^{C_a}$7 is

$\mathbf{a}_{u,v}^t \in \mathbb{R}^{C_a}$8

where $\mathbf{a}_{u,v}^t \in \mathbb{R}^{C_a}$9 is archive-based behavioral novelty and $\mathbf{d}_{u,v}^t \in \mathbb{R}^{C_d}$0 is contemporary visual diversity (Berdica et al., 13 Apr 2026). The objective is intentionally non-stationary: it rewards worlds that differ both from historical behaviors and from the current population.

For historical novelty, each rollout is summarized by a handcrafted ecological descriptor. At each time step the method computes species fractions across the $\mathbf{d}_{u,v}^t \in \mathbb{R}^{C_d}$1 agents plus environment and extracts five summary statistics: mean occupancy $\mathbf{d}_{u,v}^t \in \mathbb{R}^{C_d}$2, temporal standard deviation $\mathbf{d}_{u,v}^t \in \mathbb{R}^{C_d}$3, mean frame-to-frame occupancy change $\mathbf{d}_{u,v}^t \in \mathbb{R}^{C_d}$4, winner-map entropy $\mathbf{d}_{u,v}^t \in \mathbb{R}^{C_d}$5, and mean absolute alive-mass change $\mathbf{d}_{u,v}^t \in \mathbb{R}^{C_d}$6. These are concatenated and $\mathbf{d}_{u,v}^t \in \mathbb{R}^{C_d}$7-normalized into

$\mathbf{d}_{u,v}^t \in \mathbb{R}^{C_d}$8

A bounded FIFO archive $\mathbf{d}_{u,v}^t \in \mathbb{R}^{C_d}$9 stores past descriptors. After each meta-iteration, the top-$\mathbf{h}_{u,v}^t \in \mathbb{R}^{C_h}$0 descriptors are added, with $\mathbf{h}_{u,v}^t \in \mathbb{R}^{C_h}$1. The novelty term is the $\mathbf{h}_{u,v}^t \in \mathbb{R}^{C_h}$2-nearest-neighbor distance in archive space,

$\mathbf{h}_{u,v}^t \in \mathbb{R}^{C_h}$3

with $\mathbf{h}_{u,v}^t \in \mathbb{R}^{C_h}$4.

For contemporary diversity, the method uses a frozen DINOv2 embedding. For world $\mathbf{h}_{u,v}^t \in \mathbb{R}^{C_h}$5 at sampled frame $\mathbf{h}_{u,v}^t \in \mathbb{R}^{C_h}$6,

$\mathbf{h}_{u,v}^t \in \mathbb{R}^{C_h}$7

and the pairwise cosine distance is

$\mathbf{h}_{u,v}^t \in \mathbb{R}^{C_h}$8

To reduce outlier effects, the method takes the median distance to other worlds at each timestep and averages over the rollout,

$\mathbf{h}_{u,v}^t \in \mathbb{R}^{C_h}$9

An important interpretive point is that the framework does not define an explicit loss term for monocultures or dead states in $A^t \in \mathbb{R}^{N \times H \times W},$0. Those failures are addressed implicitly through the ecological descriptor, archive novelty, visual diversity, and the competitive PD-NCA substrate. The paper therefore distinguishes between the optimization objective and diagnostic metrics.

Two diagnostics are central. Ecological Persistence is defined from per-pixel aliveness entropy $A^t \in \mathbb{R}^{N \times H \times W},$1 as

$A^t \in \mathbb{R}^{N \times H \times W},$2

High $A^t \in \mathbb{R}^{N \times H \times W},$3 indicates sustained multi-agent coexistence. Effective complexity is defined as

$A^t \in \mathbb{R}^{N \times H \times W},$4

where $A^t \in \mathbb{R}^{N \times H \times W},$5 is normalized spatial entropy and $A^t \in \mathbb{R}^{N \times H \times W},$6 is a compressibility ratio. The metric is high when the system is neither too ordered nor too random but exhibits structured, compressible variety.

5. Experimental comparisons and emergent behavioral regimes

The reported comparisons include PBT-NCA, fixed-hyperparameter PD-NCA, and random search (RS), with RS using the same compute budget and population size (Berdica et al., 13 Apr 2026). The search space covers learning rate, batch size, steps per update, softmax temperature, and per-hidden-update frequency. Search trends show selection pressure toward higher learning rates and smaller batch sizes, which increase gradient noise and help avoid collapse into stable but uninteresting equilibria.

The baseline fixed-hyperparameter PD-NCA frequently degenerates into high-entropy noise, static frames, or low-diversity dynamics. Random search can discover some regular cyclic behavior, but it does not exhibit the sustained, open-ended progression attributed to PBT-NCA. By contrast, under the composite novelty objective, both composite score and mean novelty generally improve over meta-iterations. This is presented as evidence that the population continues to produce new kinds of worlds rather than merely settling into a local optimum.

The reported effect of agent count is also specific: worlds with more NCA agents, such as 7 agents, sustain higher novelty later in training than smaller worlds. When the hyperparameter search space is expanded, the method discovers more rigid, geometric, circuit-like patterns, including gliders, spaceships, and information-propagating structures. This suggests that the framework is not confined to one morphological family but can traverse qualitatively different dynamical classes.

The emergent phenomena described in the paper include:

• Periodic waves: highly regular, coordinated wave-like behavior.
• Spore-like scattering / colonization: homogeneous groups eject cell-like clusters that colonize distant territory.
• Migrating macro-structures with active interiors: fluid, shape-shifting entities with stable outer boundaries enclosing highly active interiors.
• Spiral entities, amoebae, aliens, gliders: long-horizon morphological differentiation into spirals, amoeba-like forms, alien-like shapes, and glider-like structures.
• Trail-following / ant-like locomotion: decentralized locomotion resembling agents following learned trails.

These observations are central to the paper’s claim that competitive differentiable cellular ecologies can be driven toward prolonged lifelike self-organization when selection pressure is defined over both present diversity and historical novelty.

6. Open-endedness, edge-of-chaos interpretation, limitations, and terminological scope

PBT-NCA is explicitly linked to open-ended discovery, effective complexity, and the edge of chaos (Berdica et al., 13 Apr 2026). The reported edge-of-chaos diagnostics are:

Metric	Reported value
$A^t \in \mathbb{R}^{N \times H \times W},$7	$A^t \in \mathbb{R}^{N \times H \times W},$8
Mean entropy $A^t \in \mathbb{R}^{N \times H \times W},$9	$N$0 bits
$N$1	$N$2

These values are interpreted as evidence that the dynamics remain between frozen order and chaotic randomness. The system is therefore argued to support persistence, information transmission, self-organization, and computation-like behavior. A plausible implication is that the composite objective is functioning not only as a novelty measure but also as a regulator that keeps evolutionary search away from both dead and saturated regimes.

The principal limitations are also clearly stated. The substrate remains 2D, which simplifies thermodynamic and environmental constraints relative to real ecosystems. The use of DINOv2 may inject anthropocentric / natural-image biases into the novelty signal. The framework still depends on a designed scoring system rather than a fully autonomous open-ended evolutionary engine. Scaling to larger grids and more agents remains an open direction, alongside hardware-accelerated CA frameworks such as CAX, larger-scale evolution strategies, and co-evolving architectures, environments, and update rules.

The term PBT-NCA is specific to Population-Based Training for Petri Dish NCA. It should be distinguished from other acronym pairings in contemporary research. In battery prognostics, PBT denotes Pretrained Battery Transformer, where NCA refers to the cathode chemistry $N$3 rather than neural cellular automata (Tan et al., 18 Dec 2025). In embedded medical imaging, NCA denotes Neural Cellular Automata deployed for bleeding segmentation and monocular depth estimation on wireless capsule endoscopy hardware, but without the PD-NCA population-evolution setting (Krumb et al., 30 Apr 2025). Within artificial life, however, PBT-NCA refers specifically to the population-based meta-evolutionary framework built on PD-NCA worlds and a dual novelty objective.