Discovering Sensorimotor Agency in Cellular Automata using Diversity Search

Code to reproduce the experiments from the paper : Discovering Sensorimotor Agency in Cellular Automata using Diversity Search Gautier Hamon¹, Mayalen Etcheverry¹, Bert Wang-Chak Chan², Clément Moulin-Frier¹, and Pierre-Yves Oudeyer¹

1. Inria Bordeaux (Flowers team), 2. Google Deepmind

Companion website for videos and demo

[Colab for the main method (IMGEP search)] (https://colab.research.google.com/drive/11mYwphZ8I4aur8KuHRR1HEg6ST5TI0RW?usp=sharing)

Contact us for questions or comments: gautier.hamon@inria. fr

Instructions

Installation

Install sensorimotor_leniasearch

pip install sensorimotor_leniasearch/.

Running experiments

Experiments are under the expe folder

For example to run one seed of imgep search run :

python expe/imgep/run_experiment.py

For each expe you can change expe_config.py to change seed, path etc.

After running a search, you can run prefilter (changing the config for the parameters generated)

python expe/prefilter/run_experiment.py

Then run the stats on long rollout with

python expe/stats_from_params/run_experiment.py

Then on the generated stats run calc_categories.py to run the empirical agency and moving test

python expe/calc_categories.py

Then run the empirical base robustness test

python expe/test_robu/run_experiment.py

For the generalization tests, we provide in expe/generalization_test all tests performed in the paper (same run python expe/generalization_test/expe_name/run_experiment.py). For slurm users, we also provide in expe/generalization_test/exputils_expe code to launch several seeds on all the generalization tests. change experiment_configurations according to your need. Then run generate_experimets.sh to generate the experiment folder and run run_slurm_experiments.ssh to launch all experiments. The exputils packages used here builds upon flowersteam's packages developped by Chris Reinke.

Parameters from the original Lenia papers (handmade search)

The code to gather the parameters from the original papers is in expe/handmade_crea_json2pickle.py . It gathers parameters from https://github.com/Chakazul/Lenia/blob/master/Python/animals.json and https://github.com/Chakazul/Lenia/tree/master/Python/found

Data and reproducing the figures

Notebooks and data to reproduce the figures of the paper can be found in the data folder