Code for "Solving Graph-based Public Good Games with Tree Search and Imitation Learning"

This is the code for the paper Solving Graph-based Public Good Games with Tree Search and Imitation Learning by Victor-Alexandru Darvariu, Stephen Hailes and Mirco Musolesi, presented at NeurIPS 2021. If you use this code, please consider citing:

@inproceedings{darvariu_solving_2021,
  title = {Solving Graph-based Public Good Games with Tree Search and Imitation Learning},
  author = {Darvariu, Victor-Alexandru and Hailes, Stephen and Musolesi, Mirco},
  booktitle = {35th Conference on Neural Information Processing Systems (NeurIPS 2021)},
  year={2021},
}

License

MIT.

Prerequisites

Please ensure that you clone this repository under the relnet root directory, e.g.

git clone [email protected]:VictorDarvariu/solving-graph-pgg.git relnet

Currently tested on Linux and MacOS (specifically, CentOS 7.4.1708 and Mac OS Big Sur 11.2.3), can also be adapted to Windows through WSL. The host machine requires NVIDIA CUDA toolkit version 9.0 or above (tested with NVIDIA driver version 384.81).

Makes heavy use of Docker, see e.g. here for how to install. Tested with Docker 19.03. The use of Docker largely does away with dependency and setup headaches, making it significantly easier to reproduce the reported results.

Configuration

The Docker setup uses Unix groups to control permissions. You can reuse an existing group that you are a member of, or create a new group groupadd -g GID GNAME and add your user to it usermod -a -G GNAME MYUSERNAME.

Create a file relnet.env at the root of the project (see relnet_example.env) and adjust the paths within: this is where some data generated by the container will be stored. Also specify the group ID and name created / selected above.

Add the following lines to your .bashrc, replacing /home/john/git/relnet with the path where the repository is cloned.

export RN_SOURCE_DIR='/home/john/git/relnet'
set -a
. $RN_SOURCE_DIR/relnet.env
set +a

export PATH=$PATH:$RN_SOURCE_DIR/scripts

Make the scripts executable (e.g. chmod u+x scripts/*) the first time after cloning the repository, and run apply_permissions.sh in order to create and permission the necessary directories.

Managing the containers

Some scripts are provided for convenience. To build the containers (note, this will take a significant amount of time e.g. 2 hours, as some packages are built from source):

update_container.sh

To start them:

manage_container_gpu.sh up
manage_container.sh up

To stop them:

manage_container_gpu.sh stop
manage_container.sh stop

To purge the queue and restart the containers (useful for killing tasks that were launched):

purge_and_restart.sh

Adjusting the number of workers and threads

To take maximum advantage of your machine's capacity, you may want to tweak the number of threads for the GPU and CPU workers. This configuration is provided in projectconfig.py. Additionally, you may want to enforce certain memory limits for your workers to avoid OOM errors. This can be tweaked in docker-compose.yml and manage_container_gpu.sh.

It is also relatively straightforward to add more workers from different machines you control. For this, you will need to mount the volumes on networked-attached storage (i.e., make sure paths provided in relnet.env are network-accessible) and adjust the location of backend and queue in projectconfig.py to a network location instead of localhost. On the other machines, only start the worker container (see e.g. manage_container.sh).

Setting up graph data

Synthetic data will be automatically generated when the experiments are ran and stored to $RN_EXPERIMENT_DIR/stored_graphs.

Accessing the services

There are several services running on the manager node.

• Jupyter notebook server: http://localhost:8888
• Flower for queue statistics: http://localhost:5555
• Tensorboard (currently disabled due to its large memory footprint): http://localhost:6006
• RabbitMQ management: http://localhost:15672

The first time Jupyter is accessed it will prompt for a token to enable password configuration, it can be grabbed by running docker exec -it relnet-manager /bin/bash -c "jupyter notebook list".

Accessing experiment data and results database

Experiment data and results are stored in part as files (under your configured $RN_EXPERIMENT_DATA_DIR) as well as in a MongoDB database. To access the MongoDB database with a GUI, you can use a MongoDB client such as Robo3T and point it to http://localhost:27017.

Some functionality is provided in relnet/evaluation/storage.py to insert and retrieve data, you can use it in e.g. analysis notebooks.

Running experiments

Experiments are launched from the manager container and processed (in a parallel way) by the workers. The file relnet/evaluation/experiment_conditions.py contains the configuration for the experiments reported in the paper, but you may modify e.g. agents, objective functions, hyperparameters etc. to suit your needs.

Then, you can launch all the experiments as follows:

Part 1: Hyperparameter optimization & evaluation for all aproaches except GIL

run_part1.sh

Part 2: Data collection for GIL using the UCT algorithm

run_part2.sh

Part 3: Training & hyperparameter optimization for GIL

run_part3.sh

Monitoring experiments

• You can navigate to http://localhost:5555 for the Flower interface which shows the progress of processing tasks in the queue. You may also check logs for both manager and worker at $RN_EXPERIMENT_DATA_DIR/logs.

Reproducing the results

Jupyter notebooks are used to perform the data analysis and produce tables and figures. Navigate to http://localhost:8888, then notebooks folder.

All tables and result figures can be obtained by opening the GGNN_Evaluation.ipynb notebook, selecting the py3-relnet kernel and run all cells. Resulting .pdf figures and .tex tables can be found at $RN_EXPERIMENT_DIR/aggregate. There are additional notebooks provided for analyzing the results of hyperparameter optimization:

• GGNN_Hyperparam_Optimisation.ipynb for UCT
• GGNN_Hyperparam_Optimisation_IL.ipynb for GIL

Problems with jupyter kernel

In case the py3-relnet kernel is not found, try reinstalling the kernel by running docker exec -it -u 0 relnet-manager /bin/bash -c "source activate relnet-cenv; python -m ipykernel install --user --name relnet --display-name py3-relnet"