As One and Many: Relating Individual and Emergent Group-Level Generative Models in Active Inference

Area of Science:

• Computational Neuroscience
• Theoretical Biology
• Artificial Intelligence

Background:

• Active inference, grounded in the Free Energy Principle, offers a unified framework for behavior and self-maintenance across scales.
• Emergent group-level agents can form from collectives of individual agents if they maintain a group-level Markov blanket.
• Understanding the generative models of these emergent group agents is challenging, limiting research in multi-scale active inference.

Purpose of the Study:

• To propose a data-driven methodology for characterizing the relationship between a group-level agent's generative model and the dynamics of its constituent individual agents.
• To demonstrate this methodology using a computational cognitive modeling approach.
• To explore the implications for understanding self-organizing systems and nested active inference agents.

Main Methods:

• Utilizing computational cognitive modeling and computational psychiatry techniques.
• Simulating a collective of agents with Markov blankets on a Multi-Armed Bandit task using the ActiveInference.jl library.
• Employing sampling-based parameter estimation to infer the generative model of the group-level agent.

Main Results:

• A non-trivial relationship was identified between the generative models of individual agents and the emergent group-level agent.
• The proposed methodology successfully characterized the link between individual and collective agent models.
• The findings hold even in a simplified Multi-Armed Bandit task setting.

Conclusions:

• The developed methodology provides a novel way to study multi-scale active inference and emergent group behavior.
• This approach can be extended to analyze nested active inference agents across various spatiotemporal scales.
• Further research can apply this methodology to diverse self-organizing systems, from cellular collectives to human societies.