Responsible: Nil Geisweiller, Yeabsira Derese (iCog)
GitHub:
Papers:
Description:
MOSES is an evolutionary program generation engine, designed to breed compact, interpretable computer programs that solve complex problems. Unlike deep neural networks that function as black boxes of opaque weights, MOSES evolves transparent symbolic code capable of logical generalization. It treats the search for solutions as a meta-optimization problem, maintaining diverse subpopulations of programs (demes) to avoid local optima while iteratively refining candidates.
MOSES combines probabilistic model-building with evolutionary search via two nested loops: an outer loop that explores structural variations (recombining program trees) and an inner loop that tunes numeric parameters within those trees.
A defining characteristic is its rigorous use of Elegant Normal Form (ENF) to constrain the search space. The system converts every candidate program into a canonical, normalized representation, eliminating redundancy and drastically shrinking the combinatorial search space.
Roadmap: