Emergence of Social Reality of Emotion through a Social Allostasis Model with Dynamic Interpretants

The evolution of emotion modeling has transitioned from biologically fixed categories to dynamic, socially constructed concepts. Early models like Damasio’s somatic marker hypothesis emphasized bodily states, while Barrett’s Embodied Predictive Interoception Coding (EPIC) formalized active inference of interoception at the individual level. Recent advances incorporate symbol systems and social interaction, exemplified by Taniguchi’s collective predictive coding hypothesis. However, existing models often treat bodily regulation and social symbol sharing separately, limiting understanding of their integration in emotion construction. This gap hampers the development of comprehensive theories of social cognition and affective AI, necessitating models that unify bodily and social processes.

The core challenge is to simulate how multiple agents develop shared emotion concepts through interaction, balancing individual bodily regulation with social symbol negotiation. Specifically, how preferences over interoceptive states and symbolic meanings co-evolve without direct access to each other's internal states. Existing models lack mechanisms for preference adaptation driven by social feedback, and do not account for the dynamic semantic shifts of symbols during social exchanges. Addressing this requires a unified framework that models both bodily control and social symbol evolution, capturing the emergent nature of social emotional reality.

The key innovations include: 1) integrating active inference with symbol evolution via a shared probabilistic model; 2) introducing a preference adaptation mechanism where agents update their prior preferences based on symbol rejection, fostering mutual alignment; 3) employing a Bayesian variational inference combined with MHNG for robust symbol exchange and semantic flexibility. These components enable the simulation of social consensus formation on emotion concepts, capturing the bidirectional influence between bodily states and social symbols, a significant step beyond prior isolated models.

• �� Construct two agents with POMDP models, defining discrete states, observations, actions, and symbols. • Each agent performs variational Bayesian inference to estimate internal states and symbols, minimizing free energy. • Actions are selected to minimize expected free energy, guided by symbol interpretations. • Symbols are inferred through active inference, with the MHNG sampling the most suitable symbol for mutual agreement. • Preference parameters are dynamically updated when symbols are rejected, moving preferences toward observed outcomes predicted by accepted symbols. • Symbol interpretation parameters E are adjusted based on the inferred symbols and actions, enabling semantic shifts. • The iterative process leads to preference convergence and symbol meaning adaptation, validated through simulation of energy and temperature regulation tasks.

Simulations involved two agents controlling energy and body temperature, starting with different preferences and no shared symbols. The agents exchanged symbols via MHNG, updating their preferences and symbol interpretations over millions of iterations. Metrics included Jensen-Shannon divergence of preferences and the evolution of symbol interpretation matrices. The setup tested whether preferences converged and whether symbols adapted their meanings in response to social feedback. Hyperparameters such as learning rates and thresholds were tuned to observe stability and convergence speed. Results demonstrated preference alignment and semantic shifts, confirming the model’s capacity to simulate social emotion formation.

Preference divergence decreased from 0.4 to 0.05, indicating strong convergence. Symbols initially encoded temperature regulation actions, later shifting to sleep, as preferences aligned. The dynamic evolution of symbol meanings reflected the social negotiation process, with preference parameters stabilizing midway through iterations. These findings support the hypothesis that social consensus on emotion concepts can emerge from simple interaction rules, driven by preference updates and symbolic negotiation, without explicit predefinition of shared meanings.

The model can inform the design of socially adaptive robots and virtual agents capable of developing shared emotional understanding with humans. It also provides a foundation for affective computing systems that adapt to social contexts. Long-term, this framework could enable large-scale social AI systems that autonomously learn and negotiate emotional concepts, fostering more natural human-AI interactions and collaborative behaviors in complex social environments.

The current model relies on discrete states and simplified tasks, limiting real-world applicability. Extending to continuous, high-dimensional environments requires substantial adaptation. The simulation environment lacks real social complexity, and the model’s scalability to larger agent populations remains untested. Parameter sensitivity and convergence stability need further investigation, especially in more diverse social scenarios. Future work should incorporate deep learning for richer representations and validate with empirical social data.