Effects of priors on epistemic uncertainty in autoregressive active inference

Abstract

Intelligent agents learn to solve tasks through interactions with their environment. We study autoregressive active inference agents that learn system dynamics and plan actions to reach a goal. From a free energy functional, we derive a variational posterior over the agent’s next action, with naturally emerging epistemic and goal-driven components. We study the effect of the epistemic term on the chosen action and investigate how it depends on prior parameters.