next up previous
Next: Computable Predictions through the Up: The New AI: General Previous: Prediction Using a Universal


Super Omegas and Generalizations of Kolmogorov Complexity & Algorithmic Probability

Our recent research generalized Solomonoff's approach to the case of less restrictive nonenumerable universal priors that are still computable in the limit [50,52].

An object $X$ is formally describable if a finite amount of information completely describes $X$ and only $X$. More to the point, $X$ should be representable by a possibly infinite bitstring $x$ such that there is a finite, possibly never halting program $p$ that computes $x$ and nothing but $x$ in a way that modifies each output bit at most finitely many times; that is, each finite beginning of $x$ eventually converges and ceases to change. This constructive notion of formal describability is less restrictive than the traditional notion of computability [67], mainly because we do not insist on the existence of a halting program that computes an upper bound of the convergence time of $p$'s $n$-th output bit. Formal describability thus pushes constructivism [5,1] to the extreme, barely avoiding the nonconstructivism embodied by even less restrictive concepts of describability (compare computability in the limit [17,40,14] and $\Delta^0_n$-describability [42][31, p. 46-47]).

The traditional theory of inductive inference focuses on Turing machines with one-way write-only output tape. This leads to the universal enumerable Solomonoff-Levin (semi) measure. We introduced more general, nonenumerable, but still limit-computable measures and a natural hierarchy of generalizations of algorithmic probability and Kolmogorov complexity [50,52], suggesting that the ``true'' information content of some (possibly infinite) bitstring $x$ actually is the size of the shortest nonhalting program that converges to $x$ and nothing but $x$ on a Turing machine that can edit its previous outputs. In fact, this ``true'' content is often smaller than the traditional Kolmogorov complexity. We showed that there are Super Omegas computable in the limit yet more random than Chaitin's ``number of wisdom'' Omega [9] (which is maximally random in a weaker traditional sense), and that any approximable measure of $x$ is small for any $x$ lacking a short description.

We also showed that there is a universal cumulatively enumerable measure of $x$ based on the measure of all enumerable $y$ lexicographically greater than $x$. It is more dominant yet just as limit-computable as Solomonoff's [52]. That is, if we are interested in limit-computable universal measures, we should prefer the novel universal cumulatively enumerable measure over the traditional enumerable one. If we include in our Bayesmix such limit-computable distributions we obtain again sharp loss bounds for prediction based on the mix [50,52].

Our approach highlights differences between countable and uncountable sets. Which are the potential consequences for physics? We argue that things such as uncountable time and space and incomputable probabilities actually should not play a role in explaining the world, for lack of evidence that they are really necessary [50]. Some may feel tempted to counter this line of reasoning by pointing out that for centuries physicists have calculated with continua of real numbers, most of them incomputable. Even quantum physicists who are ready to give up the assumption of a continuous universe usually do take for granted the existence of continuous probability distributions on their discrete universes, and Stephen Hawking explicitly said: ``Although there have been suggestions that space-time may have a discrete structure I see no reason to abandon the continuum theories that have been so successful.'' Note, however, that all physicists in fact have only manipulated discrete symbols, thus generating finite, describable proofs of their results derived from enumerable axioms. That real numbers really exist in a way transcending the finite symbol strings used by everybody may be a figment of imagination -- compare Brouwer's constructive mathematics [5,1] and the Löwenheim-Skolem Theorem [32,61] which implies that any first order theory with an uncountable model such as the real numbers also has a countable model. As Kronecker put it: ``Die ganze Zahl schuf der liebe Gott, alles Übrige ist Menschenwerk'' (``God created the integers, all else is the work of man'' [6]). Kronecker greeted with scepticism Cantor's celebrated insight [7] about real numbers, mathematical objects Kronecker believed did not even exist.

Assuming our future lies among the few (countably many) describable futures, we can ignore uncountably many nondescribable ones, in particular, the random ones. Adding the relatively mild assumption that the probability distribution from which our universe is drawn is cumulatively enumerable provides a theoretical justification of the prediction that the most likely continuations of our universes are computable through short enumeration procedures. In this sense Occam's razor is just a natural by-product of a computability assumption! But what about falsifiability? The pseudorandomness of our universe might be effectively undetectable in principle, because some approximable and enumerable patterns cannot be proven to be nonrandom in recursively bounded time.

The next sections, however, will introduce additional plausible assumptions that do lead to computable optimal prediction procedures.


next up previous
Next: Computable Predictions through the Up: The New AI: General Previous: Prediction Using a Universal
Juergen Schmidhuber 2003-11-27