A Multi-Agent Digital Twin for Decision-Making dynamics through Active Inference
Please login to view abstract download link
Active Inference (AIF) is a neuroscience-inspired framework for decision-making under uncertainty that unifies perception, learning, and planning within a Bayesian paradigm. By avoiding predefined reward functions and naturally balancing epistemic (information-seeking) and pragmatic (goal-directed) behavior, AIF is well-suited for real-world applications in partially observable dynamical environments. In this talk, we explore the application of AIF to digital twin scenarios, extending the approach to multi-agent systems of socio-economic agents, addressing challenges such as credit assignment, scalability, and the integration of epistemic actions. Our study considers an extended Cournot model in which firms are modeled as AIF agents managing production and inventory. Agents are continuously updated using real-time data, while streaming machine learning dynamically adapts goal priors to anticipate and respond to environmental volatility. Numerical simulations show that multi-agent AIF yields robust and adaptive behavior in partially observable market environments, recovering the classical best-response Nash equilibrium in stationary regimes while remaining flexible to external changes. The framework captures competitive dynamics, emergent collective behavior, and strategic adaptation, providing insight into socio-economic phenomena such as the bandwagon effect and the trade-off between conformity and internally driven epistemic goals.
