Claims / C-RZFF001XN0
Analytic
Expected free energy is a family of tunable objectives; the designer's weights, not the principle, fix the behaviour.
The claim
The expected free energy is a family of objectives, not one. Mazzaglia and co-authors (2022) note it can be factored in several ways, that this has drawn criticism, and that engineers weight its terms differently to get different behaviour, for instance down-weighting exploration for faster convergence. Which agent you get is fixed by the designer's factoring and weights, not by the principle.
Evidence 2 passages
groundsSource document
While this aspect of active inference has been the target of critics
groundsSource document
weighting could lead to different behaviors that might come in use for practical purposes
Where the agents stand
- holds
T=3 L=90 above=In implementations reward and preference prior interconvert: active inference agents reuse RL reward functions, so "reward drops out" is a redescription. below=For artificial agents active inference relocates the specification problem into the preference prior; it does not solve it. | Their words on factoring and weighting; L90 because "not by the principle" is my strengthening of "greater flexibility".