Programmatic Policies in Karel the Robot

This implements the methods and experiments of Reclaiming the Source: Programmatic versus Latent Search Spaces published in ICLR 2024.

Dependencies

We recommend using conda to install the dependencies:

conda create --name prog_policies_env --file environment.yml

If conda is not available, it is also possible to install dependencies using pip on Python 3.8:

pip install -r requirements.txt

Execution

All executable scripts are located in scripts/. These are the relevant scripts used in the paper:

• run_search.py: Runs a specified search algorithm in a specified task. Includes boilerplate to output log files and checkpoints. Can be plotted with plot_results.py.
• run_search_new.py: Same as run_search.py, but uses a cleaner implementation with no boilerplate.
• behaviour_smoothness.py, same_program_rate.py, convergence_rate.py: Scripts for topology-based evaluation. Plotted with plot_behaviour_smoothness.py, plot_same_program_rate.py, plot_convergence_rate.py, respectively.

All scripts are run with fixed random seeds for reproducibility.

Parameters for reproducibility

run_search.py expects parameters in a JSON format. The parameters used for each search method in each task in the paper are in sample_args/paper/{HC|CEBS|CEM}/*.json.

Note that all parameters use search_seed=1. To reproduce the results from the paper, run each algorithm with each seed value in [1,32].