Toy Multi agent environment

Start of a toy gym environment bank for multi-agent reinforcement learning.

This repo contains for the moment a first environment:

How to install ?

git clone https://github.com/MehdiZouitine/gym_ma_toy
cd gym_ma_toy
pip install -e .

How to use it ?

import gym
import gym_ma_toy

env = gym.make('team_catcher-v0')

obs = env.reset()
done = False
while not done:
    env.render(mode='rgb_array')
    
    action = env.action_space.sample()
    obs, reward, done, info = env.step(action)
env.close()

Team catcher:

This is a map where targets are randomly placed. Some targets do not move , other move randomly (one square at the time). Agents must capture all targets. Targets are captures when there are at least two agents on an adjacent cell of a target. The environement gives a reward point for each captured target. The episode ends when there is no more target on the map. The number of agents and targets can be arbitrarily large.

There are currently 3 implemented versions:

• V0: This is the simplest environment. Targets (RED) do not move and agents (BLUE) can only move horizontally and vertically.

- **V1**: Some targets do not move (ORANGE) but some can move randomly (RED).

• V2: Some agents move horizontally/vertically (BLUE), other move diagonally (GREEN)

- **V3**: **Partially observable** (fog of war like) environment with some agents move horizontally/vertically (BLUE), other move diagonally (GREEN)

Running multiple environment in parallel

# Running 8 environment in parallel
import gym
import gym_ma_toy

env = gym.vector.make('team_catcher-v0',num_envs=8, asynchronous=True)  

Test

pytest test/

Cite the environment as:

@misc{amarl2020
 Author = {Mehdi Zouitine, Adil Zouitine, Ahmad Berjaoui},
 Title = {Toy environment set for multi-agent reinforcement learning and more},
 Year = {2020},
}

License

This project is free and open-source software licensed under the MIT license.