A developmental model for the evolution of artificial neural networks.

Area of Science:

• Computational Neuroscience
• Developmental Biology
• Artificial Intelligence

Background:

• Artificial neural networks (ANNs) typically have fixed architectures.
• Modeling neural development and decentralized growth remains a challenge.
• Inspiration from biological nervous systems can inform ANN design.

Purpose of the Study:

• To present a novel model for decentralized growth and development of ANNs.
• To explore the use of artificial chemistry for simulating neuronal interactions.
• To facilitate the evolution of complex neural network structures.

Main Methods:

• Each artificial neuron is an autonomous unit governed by genetic information and local substrate concentrations.
• A simple artificial chemistry models the interactions of chemicals and substrates.
• A distributed genetic algorithm (GA) implemented in Java supports evolutionary experiments.

Main Results:

• Engineered genomes demonstrate the capability of artificial chemistry to grow simple networks with known physiological behaviors.
• The system is designed for the evolution of complex, decentralized neural networks.
• A platform-independent, asynchronous GA enables web-based participation in evolutionary experiments.

Conclusions:

• The proposed model offers a biologically inspired approach to decentralized ANN development.
• Artificial chemistry provides a viable mechanism for simulating complex emergent behaviors in ANNs.
• The implemented GA facilitates large-scale evolutionary exploration of neural network architectures.