Discovering sensorimotor agency in cellular automata using diversity search

Area of Science:

• Artificial Life
• Computational Biology
• Machine Learning

Background:

• Artificial life (ALife) explores self-organization of life-like behaviors in simulations.
• A key challenge in cellular automata (CA) is creating environment rules for self-organizing robust agents without predefined structures.

Purpose of the Study:

• To automate the search for environment rules in CA that lead to self-organization of agent-like structures.
• To investigate the emergence of agency, perception, and action from simple rules.

Main Methods:

• Leveraged machine learning algorithms: diversity search, curriculum learning, and gradient descent.
• Systematically searched for environmental conditions within CA simulations.
• Focused on automating the discovery of self-organizing individuals.

Main Results:

• Successfully identified environmental conditions in CA that promote self-organization of basic agency.
• Observed localized structures exhibiting coherent movement, reaction to obstacles, and integrity maintenance.
• Demonstrated strong generalization capabilities to novel environments.

Conclusions:

• The developed machine learning approach enables systematic discovery of self-organizing agents in CA.
• This research opens new avenues for artificial intelligence and synthetic bioengineering.
• Highlights the potential for emergent agency and robustness from simple computational rules.