This repository contains the source code associated with the paper published at ICLR'21 (OpenReview):

Augustus Odena, Kensen Shi, David Bieber, Rishabh Singh, Charles Sutton, Hanjun Dai. BUSTLE: Bottom-Up Program Synthesis Through Learning-Guided Exploration. International Conference on Learning Representations (ICLR), 2021.

In this research project, we use a learned model to guide a bottom-up program synthesis search to efficiently synthesize spreadsheet programs.

To cite this work, you can use the following BibTeX entry:

@inproceedings{odena2021bustle,
    title={{BUSTLE}: Bottom-Up Program Synthesis Through Learning-Guided Exploration},
    author={Augustus Odena and Kensen Shi and David Bieber and Rishabh Singh and Charles Sutton and Hanjun Dai},
    booktitle={International Conference on Learning Representations},
    year={2021},
    url={https://openreview.net/forum?id=yHeg4PbFHh}
}

Most of this project is written in Java. Use the compile.sh script to download necessary .jar files and compile the code. The clean.sh script will remove the downloaded .jar files and delete compiled .class files that were produced by compile.sh.

Use the download_sygus_benchmarks.sh script to download SyGuS benchmark tasks.

We use Python for model training. If you want to train new models, you'll need Python and TensorFlow.

The trained model is provided in the models directory. Use the run_synthesis.sh script to run synthesis using the BUSTLE model. This will output results as JSON files in a new results directory.

First, generate synthetic training data with the script generate_data.sh, which will produce a JSON file in a new training_data directory. Then, run the train.sh script, which will produce model files in the models directory.

Run tests with run_tests.sh.

This is not an official Google product.