From Interaction to Faction in Evolutionary Group Dynamics

We investigate the dynamics of opinion formation in a multi-agent system operating within a two-dimensional opinion space. Similar to models where agents interact in a physical space, our agents occupy positions in an abstract opinion space, where distance represents differences in opinion. Each agent’s opinion evolves based on local interactions with neighboring agents, mediated by a shallow neural network whose parameters are optimized through an evolutionary algorithm. This configuration allows agents to dynamically adapt their opinions based on simple, local information, leading to the spontaneous emergence of distinct opinion clusters or “factions”.

We explore how environmental parameters, such as interaction radius, opinion noise, and field of view, affect the complexity and stability of emergent opinion groups. To assess the social relevance of the model, we analyze the structure of the interaction networks formed over time, checking for the presence of characteristics common to real social networks, such as clustering coefficients, modularity, and small-world behavior. Our findings suggest that the degree of non-linearity in neural network processing correlates with the diversity and resilience of opinion groups, while simpler networks yield more uniform and predictable opinion clusters.

Our work provides insights into the fundamental mechanisms driving opinion clustering and polarization in social systems and offers a novel approach to modeling opinion dynamics using evolutionary neural networks.