Multiagent Cooperation
Environment
- Multiagent particle environment: https://github.com/openai/multiagent-particle-envs
Experiments The configuring parameters are located on the top of each execution file.
- train-iql.py: IQL with VDN mixing strategy
- train-gcn.py: IQL enhanced with Graph Convolutional Network with VDN
- train-gat.py: IQL enhanced with Graph Attentional Network with VDN
- train-gat-ind.py: IQL enhanced with Graph Attentional Network with VDN without shared weights
- train-rnn-ind.py: IQL with recurrence and VDN mixing strategy
- train-dueling-dqn.py: Independent Dueling DQN with VDN mixing strategy
- train-maddpg.py: MADDPG
- train-centr-maddpg: MADDPG with one centralized critic
Buffers
- replay_buffer.py: Save state, action, adjacency_matrix, next_action, reward, done
- replay_buffer_iql.py: Save state, action, next_action, reward, done (without GNN)
- 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.