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_delta_var=True delta_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=MMM2 use_action_repr=True delta_var_loss=True construction_delta_var=True threshold=0.3 t_max=2005000 independent_p_q=False

There are four methods for building sparse graphs:

construction_delta_abs: Using the maximum utility difference (Eq. 5 in the paper)
construction_q_var: Using the variance of payoff functions (Eq. 6 in the paper)
construction_delta_var: Using the variance of utility difference functions (Eq. 7 in the paper)
construction_attention: Using the attentional observation-based approach (Eq. 9 in the paper)

By default, they are set to False. Setting True for one of them would use the corresponding method to construct sparse graphs. Setting full_graph or random_graph to True can test complete or random coordination graphs, respectively.

There are three losses for learning sparse topologies (Eq. 8 in the paper):

l1_loss: Using
q_var_loss: Using
delta_var_loss: Using

By default, they are set to False. Setting True for one of them would use the corresponding loss.

CASEC uses construction_delta_var and delta_var_loss. 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.

mansicer/cg-pymarl