Perception, Action, and Learning in Mechanical Systems via Free Energy Minimization
Please login to view abstract download link
Active inference (AIF) is a neuroscience-inspired Bayesian framework for decision-making that unifies perception, planning, and learning. It relies on a generative model interacting with a partially observable dynamical environment. Action becomes an integral part of the inference process and emerges from the minimization of expected free energy, balancing goal-directed and epistemic behavior. Within this dual objective, information gain about hidden states and model parameters directly supports utility maximization, moving beyond passive data assimilation. Epistemic actions actively steer the system toward underexplored regimes or deliberately perturb the environment to accelerate data acquisition and model refinement. We design AIF agents capable of active learning and adaptation in mechanical systems. The first application addresses scenarios in which the generative model is initially unknown and must be learned online through controllable interaction. The resulting behavior combines goal-directed, information-gathering, and curiosity-driven components, enabling adaptation to novelty without catastrophic forgetting. Online learning is complemented by offline Bayesian model reduction, ensuring an efficient balance between the complexity and accuracy of the generative model. The second application focuses on an active digital twin framework for structural health monitoring and predictive maintenance. The framework jointly optimizes structural health and maintenance costs while acquiring new data when uncertainty about key system features increases, thereby maintaining synchronization with the physical asset.
