Presentation of Evolvability
summary
Defines genomes as computable functions and fitness as an empirical performance against an ideal function under an unknown distribution, then shows evolution is equivalent to restricted learning via CSQ equivalence (statistical queries) so some classes (e.g. parity not evolvable) are provably unreachable by strict Darwinian mutation/selection. Gives a constructive evolution rule that learns sparse linear targets under smoothed input distributions (sparse representations throughout) and models gene regulation as a promoter-as-Markov-chain whose steady state is a ratio of polynomials (matrix‑tree theorem).
AI-written summary · about