Computable Predictions

Given observation $x$ and some $\epsilon > 0$, we use AS to approximate all possibly relevant $S(xy)$ within $\epsilon$ accuracy, and predict a continuation $y$ with maximal $\bar{S}_{\epsilon}(xy)$. This ensures that no $z \neq y$ can yield some $S(xz)$ significantly exceeding $S(xy)$.

That is, with $S$ comes a computable method for predicting optimally within $\epsilon$ accuracy. This contrasts with Solomonoff's noncomputable method.

Back to Speed Prior page