Presentation of Evolvability
summary
Defines a formal model where genomes encode hypothesis functions, evolution samples finite examples, and selection keeps the best empirical performer; proves low-degree polynomials are evolvable (so linear maps n→m converge efficiently under quadratic loss) by exhibiting a polynomial-time mutation algorithm that finds near‑optimal hypotheses quantified over all distributions. Notes limitations: proof relies on convex L2 loss and arbitrary Turing‑machine mutations, and asks whether more natural mutation/crossover rules or nonconvex losses still work.
AI-written summary · about