---

Updated on 2022.03.24 DI-engine-v0.3.0 (beta)

Introduction to DI-engine (beta)

DI-engine doc | 中文文档

DI-engine is a generalized decision intelligence engine. It supports various deep reinforcement learning algorithms (link):

• Most basic DRL algorithms, such as DQN, PPO, SAC, R2D2
• Multi-agent RL algorithms like QMIX, MAPPO
• Imitation learning algorithms (BC/IRL/GAIL) , such as GAIL, SQIL, Guided Cost Learning
• Exploration algorithms like HER, RND, ICM
• Offline RL algorithms: CQL, TD3BC
• Model-based RL algorithms: MBPO

DI-engine aims to standardize different RL enviroments and applications. Various training pipelines and customized decision AI applications are also supported.

• Traditional academic environments
  • DI-zoo
• Real world decision AI applications
  • DI-star: Decision AI in StarCraftII
  • DI-drive: Auto-driving platform
  • GoBigger: Multi-Agent Decision Intelligence Environment
  • DI-smartcross: Decision AI in Traffic Light Control
• General nested data lib
  • treevalue: Tree-nested data structure
  • DI-treetensor: Tree-nested PyTorch tensor Lib
• Docs and Tutorials
  • DI-engine-docs
  • awesome-model-based-RL: A curated list of awesome Model-Based RL resources

DI-engine also has some system optimization and design for efficient and robust large-scale RL training:

• DI-orchestrator: RL Kubernetes Custom Resource and Operator Lib
• DI-hpc: RL HPC OP Lib
• DI-store: RL Object Store

Have fun with exploration and exploitation.

Installation

You can simply install DI-engine from PyPI with the following command:

pip install DI-engine

If you use Anaconda or Miniconda, you can install DI-engine from conda-forge through the following command:

conda install -c opendilab di-engine

For more information about installation, you can refer to installation.

And our dockerhub repo can be found here，we prepare base image and env image with common RL environments.

• base: opendilab/ding:nightly
• atari: opendilab/ding:nightly-atari
• mujoco: opendilab/ding:nightly-mujoco
• smac: opendilab/ding:nightly-smac

The detailed documentation are hosted on doc | 中文文档.

Quick Start

3 Minutes Kickoff

3 Minutes Kickoff (colab)

3 分钟上手中文版 (kaggle)

How to migrate a new RL Env | 如何迁移一个新的强化学习环境

Bonus: Train RL agent in one line code:

ding -m serial -e cartpole -p dqn -s 0

Supporters

↳ Stargazers

↳ Forkers

Feature

Algorithm Versatility

No	Algorithm	Label	Doc and Implementation	Runnable Demo
1	DQN		DQN doc DQN中文文档 policy/dqn	python3 -u cartpole_dqn_main.py / ding -m serial -c cartpole_dqn_config.py -s 0
2	C51		policy/c51	ding -m serial -c cartpole_c51_config.py -s 0
3	QRDQN		policy/qrdqn	ding -m serial -c cartpole_qrdqn_config.py -s 0
4	IQN		policy/iqn	ding -m serial -c cartpole_iqn_config.py -s 0
5	Rainbow		policy/rainbow	ding -m serial -c cartpole_rainbow_config.py -s 0
6	SQL		policy/sql	ding -m serial -c cartpole_sql_config.py -s 0
7	R2D2		policy/r2d2	ding -m serial -c cartpole_r2d2_config.py -s 0
8	A2C		policy/a2c	ding -m serial -c cartpole_a2c_config.py -s 0
9	PPO/MAPPO		policy/ppo	python3 -u cartpole_ppo_main.py / ding -m serial_onpolicy -c cartpole_ppo_config.py -s 0
10	PPG		policy/ppg	python3 -u cartpole_ppg_main.py
11	ACER		policy/acer	ding -m serial -c cartpole_acer_config.py -s 0
12	IMPALA		policy/impala	ding -m serial -c cartpole_impala_config.py -s 0
13	DDPG/PADDPG		policy/ddpg	ding -m serial -c pendulum_ddpg_config.py -s 0
14	TD3		policy/td3	python3 -u pendulum_td3_main.py / ding -m serial -c pendulum_td3_config.py -s 0
15	D4PG		policy/d4pg	python3 -u pendulum_d4pg_config.py
16	SAC		policy/sac	ding -m serial -c pendulum_sac_config.py -s 0
17	PDQN		policy/pdqn	ding -m serial -c gym_hybrid_pdqn_config.py -s 0
18	MPDQN		policy/pdqn	ding -m serial -c gym_hybrid_mpdqn_config.py -s 0
19	HPPO		policy/ppo	ding -m serial_onpolicy -c gym_hybrid_hppo_config.py -s 0
19	QMIX		policy/qmix	ding -m serial -c smac_3s5z_qmix_config.py -s 0
20	COMA		policy/coma	ding -m serial -c smac_3s5z_coma_config.py -s 0
21	QTran		policy/qtran	ding -m serial -c smac_3s5z_qtran_config.py -s 0
22	WQMIX		policy/wqmix	ding -m serial -c smac_3s5z_wqmix_config.py -s 0
23	CollaQ		policy/collaq	ding -m serial -c smac_3s5z_collaq_config.py -s 0
24	GAIL		reward_model/gail	ding -m serial_gail -c cartpole_dqn_gail_config.py -s 0
25	SQIL		entry/sqil	ding -m serial_sqil -c cartpole_sqil_config.py -s 0
26	DQFD		policy/dqfd	ding -m serial_dqfd -c cartpole_dqfd_config.py -s 0
27	R2D3		R2D3中文文档 policy/r2d3	python3 -u pong_r2d3_r2d2expert_config.py
28	Guided Cost Learning		reward_model/guided_cost	python3 lunarlander_gcl_config.py
29	TREX		reward_model/trex	python3 mujoco_trex_main.py
30	HER		reward_model/her	python3 -u bitflip_her_dqn.py
31	RND		reward_model/rnd	python3 -u cartpole_ppo_rnd_main.py
32	ICM		ICM中文文档 reward_model/icm	python3 -u cartpole_ppo_icm_config.py
33	CQL		policy/cql	python3 -u d4rl_cql_main.py
34	TD3BC		policy/td3_bc	python3 -u mujoco_td3_bc_main.py
35	MBPO		model/template/model_based/mbpo	python3 -u sac_halfcheetah_mopo_default_config.py
36	PER		worker/replay_buffer	rainbow demo
37	GAE		rl_utils/gae	ppo demo

means discrete action space, which is only label in normal DRL algorithms (1-18)

means continuous action space, which is only label in normal DRL algorithms (1-18)

means hybrid (discrete + continuous) action space (1-18)

means distributed training (collector-learner parallel) RL algorithm

means multi-agent RL algorithm

means RL algorithm which is related to exploration and sparse reward

means Imitation Learning, including Behaviour Cloning, Inverse RL, Adversarial Structured IL

means offline RL algorithm

means model-based RL algorithm

means other sub-direction algorithm, usually as plugin-in in the whole pipeline

P.S: The .py file in Runnable Demo can be found in dizoo

Environment Versatility

No	Environment	Label	Visualization	Code and Doc Links
1	atari			code link env tutorial 环境指南
2	box2d/bipedalwalker			dizoo link 环境指南
3	box2d/lunarlander			dizoo link 环境指南
4	classic_control/cartpole			dizoo link 环境指南
5	classic_control/pendulum			dizoo link 环境指南
6	competitive_rl			dizoo link 环境指南
7	gfootball			dizoo link 环境指南
8	minigrid			dizoo link 环境指南
9	mujoco			dizoo link 环境指南
10	PettingZoo			dizoo link 环境指南
11	overcooked			dizoo link env tutorial
12	procgen			dizoo link 环境指南
13	pybullet			dizoo link 环境指南
14	smac			dizoo link 环境指南
15	d4rl			dizoo link 环境指南
16	league_demo			dizoo link
17	pomdp atari			dizoo link
18	bsuite			dizoo link env tutorial
19	ImageNet			dizoo link 环境指南
20	slime_volleyball			dizoo link env tutorial 环境指南
21	gym_hybrid			dizoo link 环境指南
22	GoBigger			opendilab link env tutorial 环境指南
23	gym_soccer			dizoo link 环境指南
24	multiagent_mujoco			dizoo link 环境指南
25	bitflip			dizoo link 环境指南

means discrete action space

means continuous action space

means hybrid (discrete + continuous) action space

means multi-agent RL environment

means environment which is related to exploration and sparse reward

means offline RL environment

means Imitation Learning or Supervised Learning Dataset

means environment that allows agent VS agent battle

P.S. some enviroments in Atari, such as MontezumaRevenge, are also sparse reward type

Feedback and Contribution

• File an issue on Github
• Open or participate in our forum
• Discuss on DI-engine slack communication channel or QQ group (700157520)
• Contributes to our future plan Roadmap

We appreciate all the feedbacks and contributions to improve DI-engine, both algorithms and system designs. And CONTRIBUTING.md offers some necessary information.

Citation

@misc{ding,
    title={{DI-engine: OpenDILab} Decision Intelligence Engine},
    author={DI-engine Contributors},
    publisher = {GitHub},
    howpublished = {\url{https://github.com/opendilab/DI-engine}},
    year={2021},
}

License

DI-engine released under the Apache 2.0 license.