Probabilistic Tiling (Preliminary Attempt)

byDiffractor4mo7th Aug 20188 comments


Ω 3

We know a decent amount about how to get tiling in proof-based environments where the objective is for the AI to achieve a goal, or to write an AI to achieve a goal, or to write an AI that writes an AI that achieves a goal, and so on, as long as the chain of deferral is finite. The probabilistic setting is far less explored. This post will outline some necessary conditions to achieve tiling of an expected utility maximizer in fully general environments. Some of these conditions are fulfilled by a logical inductor in the limit. It's entirely possible that there are serious problems with this proof, so I'm hoping that it gets a good kicking.