A Dynamic Bayesian Model of Homeostatic Control

This paper shows how a planning as inference framework with discrete latent states can be used to implement homeostatic control by providing an agent with multivariate autonomic set points as goals. Before receiving these goals the agent navigates according to the ‘Prior Dynamics’ which embody a cognitive map of the environment. Given the goals, optimal value functions are implicitly computed using a forward and backward message passing algorithm, which is then used to construct the ‘Posterior Dynamics’. We propose that this formalism provides a useful description of computations in the mammalian Hippocampus.