Proof of Concept Code Generation using Formal Fields applied to the ARC Challenge
We will update this project with links to:
- Formal Fields Paper
- Open Expo 2020 presentation slides
- Stay tuned!
Note that even if Jazz has an Apache license, this is GPLed. The way to go is Jazz, this is research-only material. Of course, the content of the data folder belongs to @fchollet, is under Apache License 2.0 and can be found here.
This is a PoC of the Jazz platform in Python to research automated code generation. The classes with a Thelonious icon are Jazz classes with the same name (simplified for the PoC)
- Background: cyan (root class), ivory (class with parent), gray (type of Container), pink (related Jazz class not implemented)
- label: (P) (key property of the class it points to), (B) (Bebop definition), (!) (not implemented), (M) (main)
- style: dashed (parent of the class it points to), dotted (remark), Thelonious icon (Jazz class)
The code has docstrings in every class and method. This project is a PoC, the main project, with maintained documentation, is Jazz.
This is research code. The use is doing experiments and exploring the results. The class Search has most utilities, just instance a
Search object and do some calls. The script run_experiment is an example of this.
The folder /experiments has results for some experimental runs. This includes a JSON with many details about the search. The class
Search has a method show_found_code() to parse the json files and extract the code snippets.
When the experiments were run many snippets now in the code base did not exist. If you run again, you will find very few new discoveries since they were discovered in the experiments and the DSL only covers a small portion (150/800??) of the whole set of problems. Basically, almost all fruits have been reaped. To reproduce the original conditions, you should remove the snippets that were unknown at the time from the code base. You can do that by parsing the json files and see what was known/unknown at the time.