multi-agent-rlenv 3.5.5

A strongly typed Multi-Agent Reinforcement Learning framework

marlenv - A unified framework for muti-agent reinforcement learning

Documentation: https://yamoling.github.io/multi-agent-rlenv

The objective of marlenv is to provide a common (typed) interface for many different reinforcement learning environments.

As such, marlenv provides high level abstractions of RL concepts such as Observations or Transitions that are commonly represented as mere (confusing) lists or tuples.

Installation

Install with you preferred package manager (uv, pip, poetry, ...):

$ pip install marlenv[all] # Enable all features
$ pip install marlenv      # Basic installation

There are multiple optional dependencies if you want to support specific libraries and environments. Available options are:

• smac for StarCraft II environments
• gym for OpenAI Gym environments
• pettingzoo for PettingZoo environments
• overcooked for Overcooked environments

Install them with:

$ pip install marlenv[smac] # Install SMAC
$ pip install marlenv[gym,smac]  # Install Gym & smac support

Using marlenv with existing libraries

marlenv unifies multiple popular libraries under a single interface. Namely, marlenv supports smac, gymnasium and pettingzoo.

import marlenv

# You can instanciate gymnasium environments directly via their registry ID
gym_env = marlenv.make("CartPole-v1", seed=25)

# You can seemlessly instanciate a SMAC environment and directly pass your required arguments
from marlenv.adapters import SMAC
smac_env = env2 = SMAC("3m", debug=True, difficulty="9")

# pettingzoo is also supported
from pettingzoo.sisl import pursuit_v4
from marlenv.adapters import PettingZoo
pz_env = PettingZoo(pursuit_v4.parallel_env())

Designing custom environments

You can create your own custom environment by inheriting from the RLEnv class. The below example illustrates a gridworld with a discrete action space. Note that other methods such as step or render must also be implemented.

import numpy as np
from marlenv import RLEnv, DiscreteActionSpace, Observation

N_AGENTS = 3
N_ACTIONS = 5

class CustomEnv(MARLEnv[DiscreteActionSpace]):
    def __init__(self, width: int, height: int):
        super().__init__(
            action_space=DiscreteActionSpace(N_AGENTS, N_ACTIONS),
            observation_shape=(height, width),
            state_shape=(1,),
        )
        self.time = 0

    def reset(self) -> Observation:
        self.time = 0
        ...
        return obs

    def get_state(self):
        return np.array([self.time])

Useful wrappers

marlenv comes with multiple common environment wrappers, check the documentation for a complete list. The preferred way of using the wrappers is through a marlenv.Builder. The below example shows how to add a time limit (in number of steps) and an agent id to the observations of a SMAC environment.

from marlenv import Builder
from marlenv.adapters import SMAC

env = Builder(SMAC("3m")).agent_id().time_limit(20).build()
print(env.extras_shape) # -> (4, ) because there are 3 agents and the time counter

Related projects

• MARL: Collection of multi-agent reinforcement learning algorithms based on marlenv https://github.com/yamoling/marl
• Laser Learning Environment: a multi-agent gridworld that leverages marlenv's capabilities https://pypi.org/project/laser-learning-environment/