Recall the definition of AIXI: A will try to infer a simple program which takes A’s outputs as input and provides A’s inputs as output, and then choose utility maximizing actions with respect to that program.
Uh—what? That seems to be an unrecognisable version of AIXI. Just to start with, it uses all possible programs consistent with its observations to date.
If model A is substantially simpler than model B, its contribution to the expected utility calculation dominates model B’s. Of course you won’t concentrate all the mass on one model, but it is quite common to concentrate the mass on a class of nearly equivalent models.
I was giving an intuitive explanation, which seems to map pretty clearly onto the formal definition. I can be more explicit if you want, but I suspect you don’t actually find it unrecognizable.
Uh—what? That seems to be an unrecognisable version of AIXI. Just to start with, it uses all possible programs consistent with its observations to date.
If model A is substantially simpler than model B, its contribution to the expected utility calculation dominates model B’s. Of course you won’t concentrate all the mass on one model, but it is quite common to concentrate the mass on a class of nearly equivalent models.
I was giving an intuitive explanation, which seems to map pretty clearly onto the formal definition. I can be more explicit if you want, but I suspect you don’t actually find it unrecognizable.