Today’s Talk

We propose a framework to optimize hierarchical, variable-length objects under a black-box performance metric, with the ability to incorporate in situ constraints. The framework uses deep learning via a simple idea: use a large model to search the space of small models. Specifically, we use a neural network (NN) to emit a distribution over tractable discrete objects and employ a novel risk-seeking policy gradient to train the NN to generate higher scoring objects. As an application, we discuss the problem of symbolic regression, i.e., discovering the underlying mathematical expression that fits a dataset. The resulting algorithm outperforms several baseline methods in its ability to exactly recover symbolic expressions on a series of benchmark problems, both with and without added noise. We also discuss the

hybridization of our framework with other modern and classical approaches in symbolic regression, including evolutionary algorithms and linear methods. Finally, we present applications in the reinforcement learning setting, where we use our framework to discover simple and interpretable control policies. The discovered symbolic policies outperform highly tuned NN-based policies, with the additional benefit of being interpretable and verifiable.

Mikel Lanajuela, PhD

Mikel Landajuela is a machine learning researcher at Lawrence Livermore National Laboratory (LLNL) in Computational Engineering Directorate. He completed his Ph.D. at Université Pierre et Marie Curie and Inria, for which he was awarded the SMAI-GAMNI award 2017 (French Society of Industrial and Applied Mathematics) for the best thesis in numerical methods for the mechanical and engineering sciences. He has hold postdoctoral appointments at Politecnico di Milano in Italy and LLNL. Recently, his work in symbolic optimization using deep learning has led to several publications in top-tier machine learning conferences and several recognitions, including the “Interpretable Symbolic Regression for Data Science” competition-award at GECCO 2022.