Robust multi-cellular developmental design

Alexandre Devert, Nicolas Bredèche, Marc Schoenauer

Introduction

In the field of Evolutionary Design, researchers often struggle with the "genotype-to-phenotype" mapping. While direct representations work for small problems, they lack scalability. Indirect representations, inspired by biological development, are more promising but face significant hurdles: the "halting problem" (knowing when growth is finished), robustness to perturbations, and the challenge of coordinating local cell actions to form global structures.

This paper presents a continuous model for multi-cellular development where individual cells, controlled by evolved neural networks, communicate locally to "grow" into complex 2D patterns. By evolving the system to reach a stable equilibrium rather than stopping after a fixed time, the researchers produced digital organisms with extraordinary self-healing abilities, capable of perfectly reconstructing themselves after severe perturbations.

Methods

The authors use a 2D grid of identical cells. Each cell contains a Neural Network controller, specifically a Discrete Time Recurrent Neural Network (DTRNN).

The DTRNN controller architecture and local communication scheme between neighboring cells.

Cells do not know their global $(x, y)$ coordinates; they only receive "chemical" signals from their four immediate neighbors (Von Neumann neighborhood). The neural network takes these neighbor chemicals as inputs and produces two types of outputs: the cell's own state (its color/differentiation) and the chemicals it will send to its neighbors in the next time step.

The researchers used the NEAT (NeuroEvolution of Augmenting Topologies) algorithm to evolve both the structure and the weights of these neural controllers. This allows the complexity of the "brain" inside each cell to grow only as much as the task requires. Instead of a fixed timer, development stops when the system's "energy" (the sum of all neuron activations) stabilizes. If a system fails to stabilize within 1024 steps, it is deemed non-viable.

Datasets and Evaluation

The model was evaluated using the "Flag Problem," a classic benchmark in developmental robotics. The goal is to get a grid of cells to match a specific 2D target image. The authors tested four 32x32 grayscale patterns: 2-bands, 3-bands, disc, and half-discs.

The four target 32x32 grayscale patterns used to evaluate the developmental model.

They compared their developmental models against a direct regression model ($f(x, y) = z$), where a neural network is simply told the coordinates and asked to output the correct color.

Key Insights: Stability as a Global Attractor

By forcing evolution to find "fixed points," the researchers essentially evolved global attractors. The final pattern isn't just a result of a sequence of steps; it's a stable state that the physics of the system "wants" to be in.

This leads to implicit self-healing: because the target pattern is a stable equilibrium, the system is inherently robust. If perturbed by noise, the local rules evolved to maintain the fixed point naturally drive the system back to the correct global pattern. Models using recurrent neural networks and multiple chemicals generally performed better on complex, non-linear shapes compared to simpler feedforward models.

Results and Performance

On simpler patterns like the 3-bands flag, the developmental model performed almost as well as the "informed" regression model. The standout result was self-healing excellence: when subjected to massive Gaussian noise perturbing every neuron in every cell, 100% of the feedforward controllers and 75% of the recurrent controllers returned to their exact original state. Most organisms reached stability very quickly, often in fewer than 50 iterations, despite being allowed up to 1024.

Limitations and Future Work

The model struggled with circular patterns (the disc and half-discs) due to the "x-y bias" of the grid and the local communication method, which makes diagonal coordination difficult. Additionally, evolving these systems was computationally expensive. Future work aims to scale these findings to 3D structures or mechanical designs, such as evolving a robot's body and brain simultaneously. The precise theoretical reason why these specific fixed points become such strong global attractors remains a subject for further study.

Summary updated

BibTeX
@inproceedings{devert2007robust,
  author = {Alexandre Devert and Nicolas Bredèche and Marc Schoenauer},
  title = {{Robust multi-cellular developmental design}},
  booktitle = {GECCO},
  year = {2007},
  doi = {10.1145/1276958.1277156},
}
Figure