Planning

Active inference plans with expected free energy, but message passing minimises a Bethe free energy. This talk connects the two: it adds an information constraint to the Bethe free energy, which brings back the drive to explore. The constrained policy equals the expected free energy policy at a particular multiplier value. That multiplier is solved rather than tuned, so exploration switches off once the agent has enough information.

This talk presents a principled mathematical framework for deriving information-seeking behavior in intelligent agents, grounded in Bayesian inference and the free energy principle.

We present the design of an autoregressive active inference agent in the form of message passing on a factor graph.

Gaussian approximations to nonlinear observation functions hat are sensitive to curvature, such as a second-order Taylor approximation, produce a state-dependent ambiguity term in expected free energy minimization.