A collection of multi-agent OpenAI gym environments
Switch branches/tags
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
examples
mgym
.gitignore
README.md
setup.py
snake.jpg

README.md

Multi-agent gym environments

snakeclip

This repository has a collection of multi-agent OpenAI gym environments.

DISCLAIMER: This project is in its early stages --- it is still a work in progress.

Dependencies

  • gym
  • numpy

Installation

git clone https://github.com/cjm715/mgym.git
cd mgym/
pip install -e .

Environments

Examples

import gym
import mgym
import random

env = gym.make('TicTacToe-v0')
fullobs = env.reset()
while True:
    print('Player O ') if fullobs[0] else print('Player X')
    a = random.choice(env.get_available_actions())
    fullobs,rewards,done,_ = env.step(a)
    env.render()
    if done:
        break
import gym
import mgym

env = gym.make('MatchingPennies-v0')
env.reset(3)
while True:
    a = env.action_space.sample()
    _,r,done,_ = env.step(a)
    env.render()
    if done:
        break

See further examples in mgym/examples/examples.ipynb.

How are multi-agent environments different than single-agent environments?

When dealing with multiple agents, the environment must communicate which agent(s) can act at each time step. This information must be incorporated into observation space. Conversely, the environment must know which agents are performing actions. Therefore this must be communicated in the action passed to the environment. The form of the API used for passing this information depends on the type of game. The two types are

  • one-at-a-time play (like TicTacToe, Go, Monopoly, etc) or
  • simultaneous play (like Soccer, Basketball, Rock-Paper-Scissors, etc).

In the TicTacToe example above, this is an instance of one-at-a-time play. The fullobs is a tuple (next_agent, obs). The variable next_agent indicates which agent will act next. obs is the typical observation of the environment state. The action a is also a tuple given by a = (acting_agent, action) where the acting_agent is the agent acting with the action given by variable action.

Testing

To run tests, install pytest with pip install pytest and run python -m pytest