Explaining the Neuroevolution of Fighting Creatures Through Virtual fMRI.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Evolutionary Computation

Background:

• Deep learning has spurred interest in artificial neural networks (ANNs), yet general artificial intelligence remains elusive.
• Detecting emergent cognitive capabilities in complex ANNs is a significant challenge.

Purpose of the Study:

• To develop a method for detecting emergent cognitive capabilities in artificial neural networks.
• To investigate the relationship between neural structure, morphology, and cognitive function through co-evolution.

Main Methods:

• Utilized co-evolving creatures with mutable morphology (cubic splines) and neural structure (ESHyperNEAT).
• Integrated low-level sensory inputs (touch, vision, hearing) to guide creature actions.
• Developed a modular abstraction of the neural network to map neuron functions to stimuli.

Main Results:

• Discovered functional mappings between neurons and stimuli, enabling high-level abstraction of neural processes.
• Observed emergent dynamics in solitary and co-evolved creatures, including range-finding and structured communication.
• Demonstrated that modular ANN abstraction simplifies the analysis of neural implementation of behaviors.

Conclusions:

• The proposed methodology effectively detects emergent cognitive capabilities in evolving artificial agents.
• Modular abstraction of artificial neural networks is crucial for understanding complex behaviors and neural implementation.
• This approach offers a pathway toward more sophisticated artificial intelligence by analyzing emergent properties.