When I stumbled on Friston’s free energy principle (FEP) and active inference it sounded important and profound. It basically states that to predict the future, we have to create it. Using a Bayesian framework, one can introduce a prior expectation of preferred behavior, and then the agent has to resolve the difference between its expectations and reality by either updating the believes or by acting on reality.

However, the papers are very hard to read, and use a lot of math from statistical mechanics. I cannot completely put it in a reinforcement learning perspective, it feels like it does not make practical predictions. Have there been practical applications of the free energy principle that demonstrates its value?