papoudakis / casec-maco-benchmark

Codes accompanying the paper "Context-Aware Sparse Deep Coordination Graphs (https://arxiv.org/abs/2106.02886).

CASEC: Context-Aware Sparse Deep Coordination Graphs

MACO: Multi-Agent Coordination benchmark

This codebase is based on PyMARL and SMAC and contains the implementation of the Multi-Agent COordination (MACO) benchmark and CASEC algorithm.

Run an experiment

Tasks in the MACO benchmark can be found in src/envs. To run experiments on the MACO benchmark:

python src/main.py --config=casec --env-config=hallway with threshold=0.5 t_max=1050000 use_action_repr=False construction_q_var=True q_var_loss=True independent_p_q=True

To run experiments on the SMAC benchmark:

python src/main.py --config=casec --env-config=sc2 with env_args.map_name=5m_vs_6m use_action_repr=True q_var_loss=True construction_q_var=True threshold=0.3 t_max=2005000 independent_p_q=False

By default, CASEC uses construction_q_var (Eq. 4 in the paper) and q_var_loss (Eq. 8 in the paper). We also provide other methods for building sparse graphs and losses for learning sparse topologies. You can refer to Appendix B.2 for more details.

• construction_delta_abs: Eq. 30 in the paper
• construction_delta_var: Eq. 32 in the paper
• construction_attention: Eq. 35 in the paper
• l1_loss: Eq. 33 in the paper
• delta_var_loss: Eq. 34 in the paper

Setting True for one of them would use the corresponding method and loss, respectively. The config files act as defaults for an algorithm or environment. They are all located in src/config. --config refers to the config files in src/config/algs. --env-config refers to the config files in src/config/envs. All results will be stored in the Results folder. The previous config files used for the SMAC Beta have the suffix _beta.

Installation instructions

Build the Dockerfile using:

cd docker
bash build.sh

Set up StarCraft II and SMAC:

bash install_sc2.sh

This will download SC2 into the 3rdparty folder and copy the maps necessary to run over.

The requirements.txt file can be used to install the necessary packages into a virtual environment.

Saving and loading learnt models

Saving models

You can save the learnt models to disk by setting save_model = True, which is set to False by default. The frequency of saving models can be adjusted using save_model_interval configuration. Models will be saved in the result directory, under the folder named models. The directory corresponding to each run will contain models saved throughout the training process, each of which is named by the number of timesteps passed since the learning process starts.

Loading models

Learnt models can be loaded using the checkpoint_path parameter, after which the learning will proceed from the corresponding timestep.

Watching StarCraft II replays

save_replay option allows saving replays of models which are loaded using checkpoint_path. Once the model is successfully loaded, test_nepisode number of episodes are run on the test mode and a .SC2Replay file is saved in the Replay directory of StarCraft II. Please make sure to use the episode runner if you wish to save a replay, i.e., runner=episode. The name of the saved replay file starts with the given env_args.save_replay_prefix (map_name if empty), followed by the current timestamp.

The saved replays can be watched by double-clicking on them or using the following command:

python -m pysc2.bin.play --norender --rgb_minimap_size 0 --replay NAME.SC2Replay

Note: Replays cannot be watched using the Linux version of StarCraft II. Please use either the Mac or Windows version of the StarCraft II client.