MARL

MARL is a high-level multi-agent reinforcement learning library, written in Python.

Project doc : [DOC]

Installation

git clone https://github.com/blavad/marl.git
cd marl
pip install -e .

Implemented algorithms

Single-agent algorithms

Q-learning	DQN	Actor-Critic	DDPG	TD3
✔️	✔️	✔️	✔️	❌

Multi-agent algorithms

minimaxQ	PHC	JAL	MAAC	MADDPG
✔️	✔️	❌	✔️	✔️

Examples

Check existing methods

import marl

# Check available agents
print("\n| Agents\t\t", list(marl.agent.available()))

# Check available agents
print("\n| Policies\t\t", list(marl.policy.available()))

# Check available agents
print("\n| Models\t\t", list(marl.model.available()))

# Check available exploration process
print("\n| Expl. Processes\t", list(marl.exploration.available()))

# Check available experience memory
print("\n| Experience Memory\t", list(marl.experience.available()))

Train a single agent with DQN algorithm

import marl
from marl.agent import DQNAgent
from marl.model.nn import MlpNet

import gym

env = gym.make("LunarLander-v2")

obs_s = env.observation_space
act_s = env.action_space

mlp_model = MlpNet(8,4, hidden_size=[64, 32])

dqn_agent = DQNAgent(mlp_model, obs_s, act_s, experience="ReplayMemory-5000", exploration="EpsGreedy", lr=0.001, name="DQN-LunarLander")

# Train the agent for 100 000 timesteps
dqn_agent.learn(env, nb_timesteps=100000)

# Test the agent for 10 episodes
dqn_agent.test(env, nb_episodes=10)

Train two agents with Minimax-Q algorithm

import marl
from marl import MARL
from marl.agent import MinimaxQAgent
from marl.exploration import EpsGreedy

from soccer import DiscreteSoccerEnv
# Environment available here "https://github.com/blavad/soccer"
env = DiscreteSoccerEnv(nb_pl_team1=1, nb_pl_team2=1)

obs_s = env.observation_space
act_s = env.action_space

# Custom exploration process
expl1 = EpsGreedy(eps_deb=1.,eps_fin=.3)
expl2 = EpsGreedy(eps_deb=1.,eps_fin=.3)

# Create two minimax-Q agents
q_agent1 = MinimaxQAgent(obs_s, act_s, act_s, exploration=expl1, gamma=0.9, lr=0.001, name="SoccerJ1")
q_agent2 = MinimaxQAgent(obs_s, act_s, act_s, exploration=expl2, gamma=0.9, lr=0.001, name="SoccerJ2")

# Create the trainable multi-agent system
mas = MARL(agents_list=[q_agent1, q_agent2])

# Assign MAS to each agent
q_agent1.set_mas(mas)
q_agent2.set_mas(mas)

# Train the agent for 100 000 timesteps
mas.learn(env, nb_timesteps=100000)

# Test the agents for 10 episodes
mas.test(env, nb_episodes=10, time_laps=0.5)