Installing Dependencies

To run the code you first need to install the required dependencies and apply our OpenSpiel patch. To install most of the dependencies, a Conda environment.yml is provided. The following steps are required to install and patch OpenSpiel.

1. Clone the OpenSpiel GitHub repository.
2. Run the install.sh to install OpenSpiel's dependencies.
3. Apply our OpenSpiel patch by running git am <path-to-project>/open_spiel_patches/open_spiel.patch from the root of the OpenSpiel repository.
4. Create a build directory and build OpenSpiel by running

   mkdir build && cd build
   cmake -DPython3_EXECUTABLE=$(which python3) -DCMAKE_CXX_COMPILER=$(which clang++) ../open_spiel
   make -j12

5. Add the built OpenSpiel to your Python PATH using

   export PYTHONPATH=$PYTHONPATH:/<path_to_open_spiel>
   export PYTHONPATH=$PYTHONPATH:/<path_to_open_spiel>/build/python

Experiments

Algorithms

In this repository, we have population-dependent versions of D-MOMD and D-AFP. These algorithms, can be run with full observability of the population or knowledge of the population of the agent's neighbour (partial observability). We also developed a population-dependent algorithm that takes into account the previous history of the agent to solve MFGs.

	Algorithm	basic id	fully observable id	partially observable id
D-AFP	Deep Average Fictitious Play	dafp	-	-
D-MOMD	Deep Munchausen Online Mirror Descent	dmomd	-	-
D-AFPP	Population-dependent D-AFP	-	fo-dafp	po-dafp
D-MOMDP	Population-dependent D-MOMD	-	fo-dmomd	po-dmomd
RD-MOMD	Recurrent Munchausen Online Mirror Descent	rnn-momd	fo-1rd-momd / fo-trd-momd*	po-1rd-momd / po-trd-momd*

* In 1rd-momd, the algorithm samples trajectories of length 1. In trd-momd, the algorithm samples trajectories of length $T$ were $T$ is the horizon.

Running different experiments

To run different algorithms, execute inside the root folder of the repo the following command:

python src/evaluation.py --config=<config_file_path> --algorithm=<algorithm_id> --logdir=<log_dir_path>

If the configuration file or the log directory are not specified, the code will run the experiments in the 2D Crowd Environment from OpenSpiel and it will save the logs into the runs directory.

For example, to run an experiment on D-MOMD in the Four Rooms OpenSpiel MFGs environment with partial observability, execute:

python src/evaluation.py --config=configs/4rooms_config.yml --algorithm=po-dmomd

To check the results generated, simply execute:

tensorboard --logdir=<log_dir_path>