New Philosophical Work on Solomonoff Induction

I don’t know to what ex­tent MIRI’s cur­rent re­search en­gages with Solomonoff in­duc­tion, but some of you may find re­cent work by Tom Sterken­burg to be of in­ter­est. Here’s the ab­stract of his pa­per Solomonoff Pre­dic­tion and Oc­cam’s Ra­zor:

Al­gorith­mic in­for­ma­tion the­ory gives an ideal­ised no­tion of com­press­ibil­ity that is of­ten pre­sented as an ob­jec­tive mea­sure of sim­plic­ity. It is sug­gested at times that Solomonoff pre­dic­tion, or al­gorith­mic in­for­ma­tion the­ory in a pre­dic­tive set­ting, can de­liver an ar­gu­ment to jus­tify Oc­cam’s ra­zor. This ar­ti­cle ex­pli­cates the rele­vant ar­gu­ment and, by con­vert­ing it into a Bayesian frame­work, re­veals why it has no such jus­tifi­ca­tory force. The sup­posed sim­plic­ity con­cept is bet­ter per­ceived as a spe­cific in­duc­tive as­sump­tion, the as­sump­tion of effec­tive­ness. It is this as­sump­tion that is the char­ac­ter­is­ing el­e­ment of Solomonoff pre­dic­tion and wherein its philo­soph­i­cal in­ter­est lies.