MLStar A System for Synthesis of Machine-Learning Programs

This paper introduces our auto-ML system, MLStar, which uses genetic programming to create scikit-learn and Keras-based Python programs to perform supervised learning. MLStar leverages our own genetic programming system (GPStar4) and provides a greater search space compared to traditional genetic programming frameworks. Key elements that enable MLStar's performance include representing individuals as Directed Acyclic Graphs (DAGs), a rich type system to shape the kinds of graphs generated, novel genetic operators which work on the DAG structure, and advanced hyperparameter tuning via the Optuna hyperparameter optimization framework. MLStar also offers multiobjective fitnesses and a variety of complex population types. We show that MLStar performs favorably to several other auto-ML frameworks on benchmark tests. We also demonstrate that MLStar is capable of competitive solutions even when running with computationally expensive features disabled.