Evolved Neural Cellular Automata (ENCA) for ARC-AGI

Evolved Neural Cellular Automata (ENCA) solves ARC-AGI tasks with NCAs evolved using CMA-ES.

Blog: enca-arc.mmzdev.com

Usage

Install Rust following instructions at rustup.rs

Linux setup

sudo apt update && sudo apt install git build-essential -y
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh && . "$HOME/.cargo/env"
git clone https://github.com/mmz-001/enca-arc
cd enca-arc/enca

Run all commands below in /enca package.

Unzip enca/data.zip containing ARC-AGI-1 and ARC-AGI-2 tasks:

unzip data.zip

Training

Use the train binary to evolve NCAs.

Typical usage:

cargo run --release --bin train -- -t ./data/v1/arc-agi_training_challenges.json -a ./data/v1/arc-agi_training_solutions.json -r runs/v1-train -s 1 -c config.json

cargo run --release --bin train -- -t ./data/v2/arc-agi_evaluation_challenges.json -a ./data/v2/arc-agi_evaluation_solutions.json -r runs/v2-eval -s 1 -c config.json

Flags:

-h, --help                   Show help
-t, --tasks-path PATH        Path to tasks JSON
-a, --solutions-path PATH    Path to solutions JSON (used for evaluation)
-r, --out-dir DIR            Output directory for this run (default: runs/<timestamp>)
-i, --id TASK_ID             Optional single task id to train/evaluate
-s, --seed SEED              Optional seed for reproducibility.
-c, --config PATH            Optional path to config.json

Visualization

Warning

Photosensitive epilepsy warning: The NCA visualizations can include rapid flashing and high-contrast transitions. Viewer discretion is advised.

After training, use the viz binary to interactively visualize execution.

Basic usage:

cargo run --release --bin viz -- -t ./data/v1/arc-agi_training_challenges.json -a ./data/v1/arc-agi_training_solutions.json -r runs/v1-train

cargo run --release --bin viz -- -t ./data/v2/arc-agi_evaluation_challenges.json -a ./data/v2/arc-agi_evaluation_solutions.json -r runs/v2-eval

Flags:

-h, --help                  Show help
-t, --tasks-path PATH       Tasks JSON file
-a, --solutions-path PATH   Solutions JSON file for evaluation
-r, --run-dir DIR           Run output directory (models and metrics)
-i, --id TASK_ID            Optional task id to open initially

Controls:

- Space: Pause/Resume simulation
- E: Toggle Train/Test split
- R: Reset current example
- D: Next example
- A: Previous example
- Shift+D / Mouse wheel up: Next task
- Shift+A / Mouse wheel down: Previous task
- Ctrl+C: Print task ID to console

Running benchmarks

Run inference benchmarks with:

cargo bench --bench bench_nca -- --save-baseline nca

Compare against a saved baseline by

cargo bench --bench bench_nca -- --baseline nca

Validation

The submission binary runs training and produces a submission.json file with:

cargo run --release --bin submission -- -p submission.json -a ./data/v1/arc-agi_training_solutions.json

Validate the submission using ground-truth solutions with:

cargo run --release --bin check -- -p submission.json -a ./data/v1/arc-agi_training_solutions.json

Testing backends

The gpu_check binary runs random NCAs on both CPU and GPU backends and checks if the results are the same:

cargo run --release --bin gpu_check

License

This project is dual-licensed under the MIT License and the Apache License 2.0; unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual-licensed as above without any additional terms or conditions.

Contributing

Contributions are welcome!

Acknowledgements

The Rust implementation of CMA-ES optimization algorithm is based on cmaes.