Basic Usage#

Initializing Environments#

The environments in MAgent2 are implemented using PettingZoo. They can be initialized by calling their env() method and passing desired parameters:

from magent2.environments import battle_v4
env = battle_v4.env(map_size=16, render_mode='human')

Interaction Workflow#

Interacting with the environment involves iterating through the agents, pulling environment information with the last() method, and sending actions through the step() method:

env.reset()
for agent in env.agent_iter():
    observation, reward, termination, truncation, info = env.last()
    action = policy(observation, agent)
    env.step(action)

Demo#

Run a demo using PettingZoo’s random_demo function which implements this workflow with random policies:

from magent2.environments import battle_v4
from pettingzoo.utils import random_demo

env = battle_v4.env(render_mode='human')
random_demo(env, render=True, episodes=1)

For more details on the API components, see the PettingZoo Basic Usage page.

Creating New Environments#

It is recommended to create MAgent2-based environments using the PettingZoo API in order to take advantage of the general multi-agent processing infrastructure and standardization available there. magent2/environments/magent_env.py contains classes to facilitate this integration. See the reference environments included in MAgent2 for some examples. See the MAgent2 API documentation for further details about what functionalities MAgent2 exposes for environment creation.