Sounds like you may enjoy Algorithmic Information Theory. In particular, Solomonoff Induction is the most information-efficient method of inducing computable r

Sounds like you may enjoy Algorithmic Information Theory. In particular, Solomonoff Induction is the most information-efficient method of inducing computable rules from observations https://en.wikipedia.org/wiki/Solomonoff%27s_theory_of_inductive_inference https://en.wikipedia.org/wiki/Solomonoff%27s_theory_of_induc... It uses Bayes rule to update hypotheses based on observations; and its initial hypothesis is the "universal prior", which assigns probabilities to observations based on how likely they are to be output by a randomly-generated computer program (more specifically, an infinite sequence of uniformly-random bits, interpreted as a self-delimiting, Turing-complete language). The universal prior is uncomputable, but has been proven to dominate all computable probability distributions. Whilst computable approximations of the universal prior, like https://en.wikipedia.org/wiki/Speed_prior https://en.wikipedia.org/wiki/Speed_prior , aren't as information-efficient, they are more time-efficient, and approach the universal prior when given more time to run (they're identical in the limit of using infinite time). This is relevant for Occam's Razor, since the universal prior itself is dominated by the shortest program that outputs an observation, i.e. the best prediction we can make about what any computable system will do next, is to look for the simplest description of (AKA shortest program outputting) everything we've seen it do so far, and make a prediction based on what that program does next. When we say the shortest program "dominates", this means that even if every longer program agreed on some different prediction, the shortest program's prediction is still more likely (it's related to the Zeno-like summation https://en.wikipedia.org/wiki/1/2_%2B_1/4_%2B_1/8_%2B_1/16_%2B_%E2%8B%AF https://en.wikipedia.org/wiki/1/2_%2B_1/4_%2B_1/8_%2B_1/16_%... ). Hence Occam's Razor is proven to be the optimal method of predicting the behaviour of any computable system. Whether or not it's a "law of nature" depends on how we define that phrase philosophically; however, I would say the Church-Turing thesis certainly counts as one, and others agree; e.g. from https://existentialtype.wordpress.com/2012/08/09/churchs-law https://existentialtype.wordpress.com/2012/08/09/churchs-law > This is all just another instance of Church’s Law, the scientific law stating that any formalism for defining computable functions will turn out to be equivalent to, say, the λ-calculus when it comes to definability of number-theoretic functions. (Ordinarily Church’s Law is called Church’s Thesis, but for reasons given in my Practical Foundations book, I prefer to give it the full status of a scientific law.) (I highly recommend that book BTW; preprints are available online) The "physical Church-Turing thesis" is a physical law akin to relativity (it states that a particular mathematical formalism corresponds to "the real world"; despite the latter having no precise definition, and hence being mathematically unprovable, but empirically testable) https://en.wikipedia.org/wiki/Church%E2%80%93Turing_thesis#Variations https://en.wikipedia.org/wiki/Church%E2%80%93Turing_thesis#V... Hence we have a physical law which tells us the universe is computable (or, if you prefer a slightly weaker statement, that no computable method can distinguish observations of a computable univese from those of an uncomputable universe, given a finite number of observations; hence even if "the real world" were uncomputable, that wouldn't have any impact on our observations) Combining these two: Occam's Razor is a provable result of a rich and rigorous mathematical theory, and applies empirically to the universe via accepted physical law. As far as testable hypotheses go, we can perform a grue/bleen experiment quite easily ( https://en.wikipedia.org/wiki/New_riddle_of_induction https://en.wikipedia.org/wiki/New_riddle_of_induction ). Here's one to try at home: - Take any existing simulation S of a physical system P which we can compare against experiment, e.g. a computational fluid dynamics simulation of our kitchen sink. - Ensure that a "prior run" of P agrees with a "prior run" of S, e.g. starting from the same initial conditions and lasting one minute. If they don't agree, go back to step 1 and alter the setup. - Alter the code of S to get a new simulation S', which behaves identically to S during a prior run (e.g. the first minute), but differently when allowed to run longer. - Try to shrink both S and S'. We can't do this exactly (since Kolmogorov Complexity is uncomputable), but we can use systematic approaches, e.g. refactorings, inlining, applying compression algorithms, etc. which should hopefully be unbiased. - Our null hypothesis is that whichever ends up shorter (S or S'), its behaviour following the "prior run" is more likely to match the real behaviour of P than the longer program (of course neither might match; but if one of them does, we'd predict it to be the shorter)