Multiagent Cooperation

Environment

1. Multiagent particle environment: https://github.com/openai/multiagent-particle-envs

Experiments The configuring parameters are located on the top of each execution file.

1. train-iql.py: IQL with VDN mixing strategy
2. train-gcn.py: IQL enhanced with Graph Convolutional Network with VDN
3. train-gat.py: IQL enhanced with Graph Attentional Network with VDN
4. train-gat-ind.py: IQL enhanced with Graph Attentional Network with VDN without shared weights
5. train-rnn-ind.py: IQL with recurrence and VDN mixing strategy
6. train-dueling-dqn.py: Independent Dueling DQN with VDN mixing strategy
7. train-maddpg.py: MADDPG
8. train-centr-maddpg: MADDPG with one centralized critic

Buffers

1. replay_buffer.py: Save state, action, adjacency_matrix, next_action, reward, done
2. replay_buffer_iql.py: Save state, action, next_action, reward, done (without GNN)
3. prioritized_replay_buffer.py

Statistics and best models are saved under the results folder. In plotting.py, we are plotting the loss per episode and the evaluation reward per episode.