. . . you can not use your 'prior knowledge' (e.g. that targets tend to be smooth, that Occam's razor usually works, etc.) to set P(f), without making additional assumptions about the applicability of that 'knowledge'. This is because that knowledge is ultimately based on only two things: your experiences since birth, and your genome's experiences in the several billion years it's been evolving.
Source: Off-Training Set Error & a priori Distinctions between learning algorithms, Santa Fe Institute, p.16
Contributed by: Zaady