All publications

evoML Yellow Paper: Evolutionary AI and Optimisation Studio

2022·arXiv
dataprep · featuresmodel searchpopulation-basedcode opt.generated modelcostaccuracy(a) the automated pipeline(b) Pareto front, not a single winner

Figure 1: The studio pipeline. Data flows through automated preparation into population-based model search (a); candidates settle onto an accuracy–cost Pareto front (b) from which deployment selects.

The yellow paper for Artemis ML (evoML): an evolutionary AI studio that automates the machine-learning lifecycle — data preparation, feature engineering, model generation, hyperparameter optimization — under explicit multi-objective search, with the generated model code itself optimized as a final stage.

Its central workflow — candidate models scored on competing objectives, Pareto fronts instead of single winners, held-out validation before anything ships — is the exact workflow Artemis later generalized from ML models to arbitrary code.

Key results

  • End-to-end automated ML with multi-objective optimization
  • Model code optimization as a first-class pipeline stage
  • The direct ancestor of Artemis's search-and-validate loop