This repository is the official implementation of Autofocused oracles for model-based optimization, to appear at NeurIPS 2020.

Notebooks for running the superconductor design experiments (e.g., Table 1 in the paper) are superconductivity_groundtruth.ipynb and superconductivity.ipynb. For the 1-D illustrative example, see toy.ipynb.

Pre-trained initial oracles for the superconductor design experiments can be found in initial_oracles. The ground-truth model is gt_dim60.model, and the initial search model is saved in init_searchmodel.npz. See supeconductivity.ipynb for how to reproduce and invoke these.

Superconductor design results are evaluated in results.ipynb, which reproduces Table 1 in the paper.