ModelRepo

Reproduce (deep) RL or Multi-Agent RL models (all of implementations are supported by gym-based (gym, particle) environments.)

Structure

lib contains environments designed for (Deep) RL and multi-agent RL tasks and dependency files, e.g, lib/ma_env is a particle environment developed by OpenAI. All algorithms listed below are implemented independently in different sub-directory.

Guides

NFSP (Neural Fictitious Self-Play)

NFSP is a framework for improving the performance of deep reinforcement learning tasks. You can get more details in 👇

• arXiv link: Deep Reinforcement Learning from Self-Play in Imperfect-Information Games

• the work of implementation is still in process ...

Multi-Agent Deep Deterministic Policy Gradient

A multi-agent deterministic policy gradient framework is proposed by Ryan Lowe and Yi Wu at 2017 which solves the non-stationary problem at training stage. Reading more by visiting 👇 arXiv link. This implementation supports gym-based multi-agent environments.

• arXiv link: Multi-Agent Actor-Critic Mixed Cooperative-Competitive Environments

• run ./scripts/run_maddpg.py to train the model, get more information about execution: python ./scripts/run_maddpg.py -h

• if you wanna try different parameters configuration, you can modify the config.py

CommNet: Learning Multiagent Communication with Backpropagation

A simple multi-agent communication framework proposed by Sainbayar Sukhbaatar, Arthur Szlam and Rob Fergus at 2016.

• arXiv link: Learning Multiagent Communication with Backpropagation

• run leaver_train.py to train the model, you can also visit KornbergFrsnel: CommNet directly to get more information

• the paper has several different playgrounds, while there has only Leaver implemented in this repo so far (maybe I will add more playgrounds, but who knows.)

DDPG (CONTINUOUS CONTROL WITH DEEP REINFORCEMENT LEARNING)

Independent Q-learning Network: (D)DQN / (D)DQN with Dueling

SAC: Soft Actor Critic

MASQ: Multiagent Soft Q-learning