sebdumancic / knorf_aaai21

This repository contains the experiments for the paper:

S. Dumancic, T. Guns, and A. Cropper: Knowledge Refactoring for Inductive Program Synthesis. AAAI '21

For any questions, contact Sebastijan

Reproducing results

To reproduce the experiments, you need to perform several steps:

1. Generate and learn play tasks

   Go to the playgol_experiments folder and follow the instructions. This is a slight modification of the Playgol, with clause deduplication disabled.

   NOTE: this step is not needed to reproduce the experiments. This step essentially produces data; you should use the same data as in the paper.

2. Refactor programs

   Go to experiments/knorf folder.

   Run python refactor_programs.py strings pncorrect "thr:8" to refactor string transformation programs (thr:8 means "use 8 threads"). To refactor lego programs, replace strings with `lego.

3. Learn programs

   Go to the experiments/knorf/runners folder. For string transfromation experiments, go further to strings folder; for lego tasks, go to lego folder.

   Run python runner.py learn-build-p "" to solve the build/target tasks.

4. Test the programs

   Run python runner.py test "" in the same folder.

5. Obtain results

   To get results in JSON format, run python runner.py results "".

   To get program sizes, replace results with size. To get the runtimes, replace results with runtime-trial.

Results with redundancies removed in a naive way

To replicate the results with naive removal of redundancies (i.e., introduce a new predicate for each redundancy in the program), follow these steps:

1. Remove redundancies

   Go to the experiments/knorf folder.

   Run python compress_redundancies.py strings pncorrect "" (use lego for Lego experiments).

2. Repeat steps 3-5 from the procedure above

   Use runner_nonredundant.py instead of runner.py