Multi-agent Communication Algs

Implemented based on Tianshou and Pettingzoo.

Tutorials

Example Scripts

On PistonBall environment

• test/test_mappo.py
• test/test_qmix.py

Algs

Comm-free baselines

• IDDPG
• MADDPG
• ISAC
• MASAC
• PPO
• MAPPO
• IQL
• QMIX
• QMIX-Attention

Comm baselines

• DIAL
• CommNet
• TarMAC
• I2C
• Intention Sharing

Simple Taxonomy

flowchart TD
    subgraph CTDE [CTDE]
        subgraph p [Parameter]
            IP([Individual Parameter])
            PS([Parameter Sharing])
            IPGI([Individual Parameter with Global Information])
        end
        subgraph c [Critic]
            IC([Individual Critic])
            JC([Joint Critic])
        end
    end
    subgraph FD [Fully Decentralized]
    end

Loading

DataFlow in MARL-Comm

We recommend A Survey of Multi-Agent Reinforcement Learning with Communication for a detailed taxonomy.

Training Schemes

Types	Sub-types
Fully Decentralized
CTDE	Individual Parameter
	Parameter Sharing
	Individual Parameter with Global Info

Logic of Tianshou in MARL

Tianshou API Overview

The figure refers to https://colab.research.google.com/drive/1MhzYXtUEfnRrlAVSB3SR83r0HA5wds2i?usp=sharing.

MARL Design Overview

flowchart
    subgraph mapolicy [MA-policy]
        p1((Agent1)) --action--> m([Manager]) -->|obs or messages| p1
        p2((Agent2)) --action--> m([Manager]) -->|obs or messages| p2
        p3((Agent3)) --action--> m([Manager]) -->|obs or messages| p3
    end    
    
    subgraph collector [Collector]
        VE(VecEnv) ==Transition==> mapolicy ==Action==> VE;
    end
    
    subgraph alg [Algorithm]
        collector ==Data==> B[(Buffer)] ==Sample==> T{Trainer} ==>|Processed Sample| mapolicy ==Info==> T
        T ==Info==> L{{Logger}}
    end

Loading

An algorithm corresponds to

• A MA-policy: interaction among agents, such as the communication
• A Buffer: what to store
• A Trainer: Update, but implemented in each agent's policy actually

Base MA components

Buffer

To facilitate the sub-types of CTDE scheme, the buffer should consider the agent dimension.

1. For the case using a single env, implement the MAReplayBuffer as a ReplayBufferManager with agent_num buffers.
2. For the case using a vectorized env, implement the VectorMAReplayBuffer as a ReplayBufferManager with agent_num $*$ env_num buffers.

For both cases, the buffer should contain sub-buffers for shared global information, e.g., the state.

PettingZoo Env Wrapper

Act as a vectorized env with agent_num envs.

Collector

AysncCollector with self._ready_env_ids initialized as [1 0 ... ] and self.data initialized to be length env_num seems suitable, accompanied with env_id in returned info.

MAPolicyManager

Maintain the centralized part inside the class.

Instructions

Install

sudo apt install swig -y
pip install 'pettingzoo[all]'
# Install my version of tianshou
git clone https://github.com/Leo-xh/tianshou.git
cd tianshou
pip install -e .