We include a non-exhaustive number of examples, showing common use-cases for Mava. We also have a Quickstart notebook that can be used to quickly create and train your first Multi-Agent System.

In Mava, we support a variety of different environments, which include PettingZoo (Repo, Paper), SMAC (Repo, Paper), 2D RoboCup, Flatland, OpenSpiel (Repo, Paper) environments, as well as a few custom environments inside Mava.

With our integration with PettingZoo, we support popular Multi-Agent environments such as SISL (Repo, Paper), MPE (Repo, Paper) and Multi-Agent Atari environments.

We include a number of systems running on continuous control tasks.

• MADDPG: a MADDPG system running on the continuous action space simple_spread MPE environment.

  • Feedforward:
    • decentralised, decentralised record agents (Example recording agents acting in the environment), decentralised scaling (Example scaling to 4 executors), decentralised custom loggers (Example using custom logging), decentralised lr scheduling (Example using lr schedule), centralised, networked (Example using a fully-connected, networked architecture), networked with custom architecture (Example using a custom, sparse, networked architecture) and state_based .
  • Recurrent
    • decentralised and state_based.

• MAD4PG: a MAD4PG system running on the continuous action space simple_spread MPE environment.

  • Feedforward
    • decentralised, centralised and state_based.
  • Recurrent
    • decentralised.

• MADDPG: a MADDPG system running on the Multiwalker environment.

  • Feedforward
    • decentralised and centralised.
  • Recurrent
    • decentralised.

• MAD4PG: a MAD4PG system running on the Multiwalker environment.

  • Feedforward
    • decentralised and decentralised record agents (Example recording agents acting in the environment).

• MAD4PG: a MAD4PG system running on the RoboCup environment.
  • Recurrent
    • state_based.

We also include a number of systems running on discrete action space environments.

• MAPPO: a MAPPO system running on the discrete action space simple_spread MPE environment.

  • Feedforward
    • decentralised and centralised.

• MADQN: a MADQN system running on the discrete action space simple_spread MPE environment.

  • Feedforward
    • decentralised, decentralised lr scheduling (Example using lr schedule), decentralised custom lr scheduling (Example using custom lr schedule) and decentralised custom epsilon decay scheduling (Example using configurable epsilon scheduling).
  • Recurrent
    • decentralised and decentralised with coms (Example using a system with communication).

• QMIX: a QMIX system running on the discrete action space simple_spread MPE environment.

  • Feedforward
    • decentralised.

• VDN: a VDN system running on the discrete action space simple_spread MPE environment.

  • Feedforward
    • decentralised.

• DIAL: a DIAL system running on the discrete action space simple_spread MPE environment.

  • Recurrent
    • decentralised.

• DIAL: a DIAL system running on the discrete custom SwitchGame environment.
  • Recurrent
    • decentralised.

• MADQN: a MADQN system running on the two-player competitive Atari Pong environment.
  • Feedforward
    • decentralised.

• MADDPG: a MADDPG system running on the Simple Speaker Listener environment.

  • Feedforward
    • decentralised.

• MADDPG: a MADDPG system running on the Simple Spread environment.

  • Feedforward
    • decentralised.

• MADQN: a MADQN system running on the SMAC environment.

  • Feedforward
    • decentralised.
  • Recurrent
    • decentralised with custom agent networks (Example using custom agent networks).

• QMIX: a QMIX system running on the SMAC environment.

  • Feedforward
    • decentralised.

• VDN: a VDN system running on the SMAC environment.

  • Feedforward
    • decentralised and decentralised record agents.

• MADQN: a MADQN system running on the OpenSpiel environment.
  • Feedforward
    • decentralised.