Consider a probabilistic model over the world which is a light-tailed probability distribution w.r.t. any physical quantity. Light-tailed means that p(X) decreases at least exponentially with X.
We do not know what physical quantities exist, and Solomonoff induction requires us to consider all computable possibilities compatible with observations so far. Any distribution p so light-tailed as to go to zero faster than every computable function must itself be uncomputable, and therefore inaccessible to Solomonoff induction. Likewise universally sub-exponentially growing utility functions.
We do not know what physical quantities exist, and Solomonoff induction requires us to consider all computable possibilities compatible with observations so far.
Yes.
Any distribution p so light-tailed as to go to zero faster than every computable function must itself be uncomputable, and therefore inaccessible to Solomonoff induction.
It doesn’t have to decrease faster than every computable function, only to decrease at least as fast as an exponential function with negative exponent.
Solomonoff induction doesn’t try to learn your utility function. Clearly, if your utility function is super-exponential, then p(X) * U(X) may diverge even if p(X) is light-tailed.
We do not know what physical quantities exist, and Solomonoff induction requires us to consider all computable possibilities compatible with observations so far. Any distribution p so light-tailed as to go to zero faster than every computable function must itself be uncomputable, and therefore inaccessible to Solomonoff induction. Likewise universally sub-exponentially growing utility functions.
Yes.
It doesn’t have to decrease faster than every computable function, only to decrease at least as fast as an exponential function with negative exponent.
Solomonoff induction doesn’t try to learn your utility function.
Clearly, if your utility function is super-exponential, then p(X) * U(X) may diverge even if p(X) is light-tailed.