This code can be used to reproduce the results in the paper On the Utility of Learning about Humans for Human-AI Coordination. Note that this repository uses a specific older commit of the overcooked_ai repository, and should not be expected to work with the current version of that repository.

To play the game with trained agents, you can use Overcooked-Demo.

For more information about the Overcooked-AI environment, check out this repo.

When cloning the repository, make sure you also clone the submodules (this implementation is linked to specific commits of the submodules, and will mostly not work with more recent ones):

$ git clone --recursive https://github.com/HumanCompatibleAI/human_aware_rl.git

If you want to clone a specific branch with its submodules, use:

$ git clone --single-branch --branch BRANCH_NAME --recursive https://github.com/HumanCompatibleAI/human_aware_rl.git

It is useful to setup a conda environment with Python 3.7:

$ conda create -n harl python=3.7
$ conda activate harl

To complete the installation, run:

               $ cd human_aware_rl
human_aware_rl $ ./install.sh

Then install tensorflow and mpi4py (the GPU or non-GPU version depending on your setup):

$ pip install tensorflow==1.13.1
$ conda install mpi4py

$ pip install tensorflow-gpu==1.13.1
$ conda install mpi4py

Note that using tensorflow-gpu will not enable to pass the DRL tests due to intrinsic randomness introduced by GPU computations. We recommend to first install tensorflow (non-GPU), run the tests, and then install tensorflow-gpu.

To verify your installation, you can try running the following command from the inner human_aware_rl folder:

python run_tests.py

Note that most of the DRL tests rely on having the exact randomness settings that were used to generate the tests (and thus will not pass on a GPU-enabled device).

On OSX, you may run into an error saying that Python must be installed as a framework. You can fix it by telling Matplotlib to use a different backend.

ppo/ (both using baselines):

• ppo.py: train one agent with PPO in Overcooked with other agent fixed

pbt/ (all using baselines):

• pbt.py: train agents with population based training in overcooked

imitation/:

• behaviour_cloning.py: simple script to perform BC on trajectory data using baselines

human/:

• process_data.py script to process human data in specific formats to be used by DRL algorithms
• data_processing_utils.py utils for the above

experiments/: folder with experiment scripts used to generate experimental results in the paper

baselines_utils.py: utility functions used for pbt.py overcooked_interactive.py: script to play Overcooked in terminal against trained agents run_tests.py: script to run all tests

To play with trained agents in the terminal, use overcooked_interactive.py. A sample command is:

python overcooked_interactive.py -t bc -r simple_bc_test_seed4

Playing requires not clicking away from the terminal window.

This requires converting the trained models to Tensorflow JS format, and visualizing with the overcooked-demo code. First install overcooked-demo and ensure it works properly.

Unfortunately, converting models requires creating a new conda environment to avoid module conflicts.

Create and activate a new conda environment:

$ conda create -n model_conversion python=3.7
$ conda activate model_conversion

Run the base setup.py (from the inner human_aware_rl) and then install tensorflowjs:

human_aware_rl $ cd human_aware_rl
human_aware_rl $ python setup.py develop
human_aware_rl $ pip install tensorflowjs==0.8.5

To convert models in the right format, use the convert_model_to_web.sh script. Example usage:

human_aware_rl $ ./convert_model_to_web.sh ppo_runs ppo_sp_simple 193

where 193 is the seed number of the DRL run.

The converted models can be found in human_aware_rl/data/web_models/ and should be transferred to the static/assets folder with the same naming as the standard models.

To play with newly trained agents, just follow the instructions in the Overcooked-Demo README.

All DRL results can be reproduced by running the .sh scripts under human_aware_rl/experiments/.

All non-DRL results can be reproduced by running cells in NeurIPS Experiments and Visualizations.ipynb.