Universal prior

A “uni­ver­sal” prior is a prob­a­bil­ity dis­tri­bu­tion that as­signs pos­i­tive prob­a­bil­ity to ev­ery think­able hy­poth­e­sis, for some rea­son­able mean­ing of “ev­ery think­able hy­poth­e­sis”. A cen­tral ex­am­ple would be Solomonoff in­duc­tion, in which the ob­ser­va­tions are a se­quence of bits, the uni­ver­sal prior is “ev­ery pos­si­ble com­puter pro­gram that gen­er­ates bit se­quences”, and ev­ery such com­puter pro­gram starts with a pos­i­tive prob­a­bil­ity.

Parents:

  • Inductive prior

    Some states of pre-ob­ser­va­tion be­lief can learn quickly; oth­ers never learn any­thing. An “in­duc­tive prior” is of the former type.