I read many of Friston’s paper on Active Inference like R20;Reinforcement Learning or Active Inference?”(at PLOSOne), “Active Inference and Learning” (at Neuroscience and Biobehavioral Reviews) and many more.
But I can’t help feel that a lot of claims he has made is a bit of a stretch. For instance, in one of his presentations on Youtube (watch?v=Y1egnoCWgUg&t=555s), he uses an example of own to motivate Active Inference and compare it to RL. However, it feels like he is just pointing out the difference between an MDP and POMDP? Also, the fact that there is no reward parameter and you need to “hack” the prior to get the desired behavior is also troubling (look at the example in the PLOSOne paper mentioned above).
I thought I will ask your opinion on his approach and the grand claims he makes that Active Inference is the way the human brain works. This is my first post. Go easy on me if I completely missed the point about Active Inference.