Evolution of neural symmetry and its coupled alignment to body plan morphology

Ben Jones, Andrea Soltoggio, Bernhard Sendhoff, Xin YaoGECCO 2011 · DOI

TL;DR

This paper explores how the evolution of an artificial organism's nervous system is intrinsically linked to its body shape. By simulating the evolution of a fish-like agent, they found that more complex motor tasks lead to more symmetrical neural structures, and that neural systems adapt to compensate for asymmetrical bodies, suggesting a deep, co-evolutionary relationship.

What Problem Does This Paper Address?

The relationship between an animal's body plan and the evolution of its nervous system is a fundamental question in biology, yet its precise nature and underlying principles remain largely unclear. While it's intuitive that body shape influences neural architecture and behavior, the extent of this influence and how it drives the co-evolution of both systems is not well understood. This paper aims to shed light on this by investigating how neural structures evolve in response to different body morphologies and motor requirements in a simulated, artificial life environment.

Methods

The researchers employed an artificial evolution approach to evolve a simulated fish-like agent. The agent's body was modeled as a series of connected segments actuated by lateral springs. A computational neural system, composed of leaky integrator neurons, controlled these springs. The key innovation lies in how the neural system's architecture (neuron positions) and connectivity were evolved. Specifically, the positions of descending neurons and head-CPG neurons were evolved, and the strength of connections between neurons was determined by their Euclidean distance. This allowed for the autonomous emergence of neural structures.

The experiments involved three scenarios:

  1. Forward locomotion: The agent needed to swim efficiently.
  2. Turning behavior: The agent had to swim and turn towards targets.
  3. Asymmetric body plan: The agent had a pre-defined curved body shape (either left or right biased) and needed to perform forward locomotion.

A measure of neural symmetry was introduced to quantify the symmetry of the descending neuron positions.

Datasets

Not discussed in the paper. The study uses a simulated environment and an artificial evolutionary process, rather than pre-existing datasets.

Key Insights and Ideas

  • Complexity breeds symmetry: The more complex the motor task (e.g., requiring turning in addition to forward motion), the more symmetrical the evolved neural structures become. This suggests that symmetry in the neural control system might be an efficient strategy for coordinating complex movements.
  • Neural systems compensate for body asymmetry: When the body plan itself is asymmetrical, the neural system evolves to adapt and compensate. Neurons and connections are arranged to counteract the body's inherent bias, allowing the agent to still achieve efficient forward movement. This highlights a remarkable ability of the nervous system to "learn" and adapt to its physical substrate.
  • Co-evolution is key: The study strongly suggests that neural architecture and body morphology do not evolve in isolation but rather in a tightly coupled, co-evolutionary manner. Each influences the other, leading to highly functional integrated systems.
  • Emergent control strategies: The research demonstrates how sophisticated control strategies, like oscillatory patterns for movement, can emerge autonomously from simple neural architectures and body mechanics, without explicit programming.

Key Results

  • Scenario (a) - Forward Motion: Evolved agents significantly improved their forward swimming ability, covering an average of 3.494 body lengths in 1600 steps, compared to 0.5898 before evolution.
  • Scenario (b) - Turning Behavior: Agents evolved the capability to turn effectively, demonstrated by visualizations of a complete U-turn and tracking a moving target in a figure-of-8 pattern.
  • Neural Symmetry: Figure 8 shows that neural symmetry evolved to greater levels (smaller values) in Scenario (b) (forward motion plus turning) compared to Scenario (a) (forward motion only).
  • Scenario (c) - Asymmetric Body: Statistical analysis showed that evolved neural architectures were biased to compensate for body asymmetry. Specifically, fewer descending connections evolved on the convex side of the body, and neurons tended to arrange themselves on the concave side. This asymmetry in neural connectivity was statistically significant (p-values of 0.04 and 0.07 for left and right biases, respectively).
  • Functional Compensation: In asymmetric bodies, the concave side's neural pathway remained intact, propagating oscillations, while the convex side's pathway was often broken, leading to tonic activation of motor neurons. This combined effect allowed the agent to maintain forward locomotion.

Limitations and Future Work

  • Simplicity of the model: The animat is a simplified, fish-inspired model. Future work could explore more complex body plans and environments.
  • Coronal plane movement: The current model focuses on movement in the sagittal plane. Extending this to include coronal plane movements could reveal further insights.
  • Energy metrics: Incorporating energy efficiency metrics could lead to the discovery of even more efficient neural structures.
  • Plastic neural connectivity: Exploring models with plastic neural connectivity, where connections can change over time, could uncover more subtle interactions between neural architectures and body morphologies.
  • HyperNEAT: The authors suggest that using more advanced neuroevolutionary techniques like HyperNEAT could allow for the encoding of more complex emergent properties like symmetry and repetition, and potentially lead to richer behaviors.

Summary updated

BibTeX
@inproceedings{jones2011evolution,
  author = {Ben Jones and Andrea Soltoggio and Bernhard Sendhoff and Xin Yao},
  title = {{Evolution of neural symmetry and its coupled alignment to body plan morphology}},
  booktitle = {GECCO},
  year = {2011},
  doi = {10.1145/2001576.2001609},
}
Figure