qun33 / rllte

Long-Term Evolution Project of Reinforcement Learning

RLLTE: Long-Term Evolution Project of Reinforcement Learning is inspired by the long-term evolution (LTE) standard project in telecommunications, which aims to track the latest research progress in reinforcement learning (RL) and provide stable and efficient baselines. In rllte, you can find everything you need in RL, such as training, evaluation, deployment, etc.

If you use rllte in your research, please cite this project like this:

@software{rllte,
  author = {Mingqi Yuan, Zequn Zhang, Yang Xu, Shihao Luo, Bo Li, Xin Jin, and Wenjun Zeng},
  title = {RLLTE: Long-Term Evolution Project of Reinforcement Learning},
  url = {https://github.com/RLE-Foundation/rllte},
  year = {2023},
}

Contents

• Contents
• Overview
• Quick Start
• Implemented Modules
• Benchmark
• API Documentation
• How To Contribute
• Acknowledgment

Overview

For the project tenet, please read Evolution Tenet.

The highlight features of rllte:

• 👨‍✈️ Large language model-empowered copilot;
• ⏱️ Latest algorithms and tricks;
• 📕 Standard and sophisticated modules for redevelopment;
• 🧱 Highly modularized design for complete decoupling of RL algorithms;
• 🚀 Optimized workflow for full hardware acceleration;
• ⚙️ Support custom environments and modules;
• 🖥️ Support multiple computing devices like GPU and NPU;
• 🛠️ Support RL model engineering deployment (TensorRT, CANN, ...);
• 💾 Large number of reusable benchmarks (See rllte-hub);

See the project structure below:

• Common: Base classes and auxiliary modules like logger and timer.

• Xploit: Modules that focus on exploitation in RL.

  • Encoder: Neural nework-based encoders for processing observations.
  • Agent: Agents for interacting and learning.
  • Storage: Storages for storing collected experiences.

• Xplore: Modules that focus on exploration in RL.

  • Augmentation: PyTorch.nn-like modules for observation augmentation.
  • Distribution: Distributions for sampling actions.
  • Reward: Intrinsic reward modules for enhancing exploration.

• Hub: Reusable datasets and models.

• Env: Packaged environments (e.g., Atari games) for fast invocation.

• Evaluation: Reasonable and reliable metrics for algorithm evaluation.

• Pre-training: Methods of pre-training in RL.

• Deployment: Methods of model deployment in RL.

For more detiled descriptions of these modules, see https://docs.rllte.dev/api

Quick Start

Installation

• Prerequisites

Currently, we recommend Python>=3.8, and user can create an virtual environment by

conda create -n rllte python=3.8

• with pip recommended

Open up a terminal and install rllte with pip:

pip install rllte-core # basic installation
pip install rllte-core[envs] # for pre-defined environments

• with git

Open up a terminal and clone the repository from GitHub with git:

git clone https://github.com/RLE-Foundation/rllte.git

After that, run the following command to install package and dependencies:

pip install -e . # basic installation
pip install -e .[envs] # for pre-defined environments

For more detailed installation instruction, see https://docs.rllte.dev/getting_started.

Start Training

On NVIDIA GPU

For example, we want to use DrQ-v2 to solve a task of DeepMind Control Suite, and it suffices to write a train.py like:

# import `env` and `agent` api
from rllte.env import make_dmc_env 
from rllte.xploit.agent import DrQv2

if __name__ == "__main__":
    device = "cuda:0"
    # create env, `eval_env` is optional
    env = make_dmc_env(env_id="cartpole_balance", device=device)
    eval_env = make_dmc_env(env_id="cartpole_balance", device=device)
    # create agent
    agent = DrQv2(env=env, 
                  eval_env=eval_env, 
                  device='cuda',
                  tag="drqv2_dmc_pixel")
    # start training
    agent.train(num_train_steps=5000)

Run train.py and you will see the following output:

On HUAWEI NPU

Similarly, if we want to train an agent on HUAWEI NPU, it suffices to replace DrQv2 with NpuDrQv2:

# import `env` and `agent` api
from rllte.env import make_dmc_env 
from rllte.xploit.agent import DrQv2

if __name__ == "__main__":
    device = "npu:0"
    # create env, `eval_env` is optional
    env = make_dmc_env(env_id="cartpole_balance", device=device)
    eval_env = make_dmc_env(env_id="cartpole_balance", device=device)
    # create agent
    agent = DrQv2(env=env, 
                  eval_env=eval_env, 
                  device='cuda',
                  tag="drqv2_dmc_pixel")
    # start training
    agent.train(num_train_steps=5000)

Then you will see the following output:

Please refer to Implemented Modules for the compatibility of NPU.

For more detailed tutorials, see https://docs.rllte.dev/tutorials.

Implemented Modules

RL Agents

Type	Module	Recurrent	Box	Discrete	MultiBinary	Multi Processing	NPU	Paper	Citations
Original	SAC	❌	✔️	❌	❌	❌	✔️	Link	5077⭐
	DDPG	❌	✔️	❌	❌	❌	✔️	Link	11819⭐
	PPO	❌	✔️	✔️	✔️	✔️	✔️	Link	11155⭐
	DAAC	❌	✔️	✔️	✔️	✔️	✔️	Link	56⭐
	IMPALA	✔️	✔️	✔️	❌	✔️	❌	Link	1219⭐
Augmented	DrQ-v2	❌	✔️	❌	❌	❌	✔️	Link	100⭐
	DrQ	❌	✔️	❌	❌	❌	✔️	Link	433⭐
	DrAC	❌	✔️	✔️	✔️	✔️	✔️	Link	29⭐

• DrQ=SAC+Augmentation, DDPG=DrQ-v2-Augmentation, DrAC=PPO+Augmentation.
• 🐌: Developing.
• NPU: Support Neural-network processing unit.
• Recurrent: Support recurrent neural network.
• Box: A N-dimensional box that containes every point in the action space.
• Discrete: A list of possible actions, where each timestep only one of the actions can be used.
• MultiBinary: A list of possible actions, where each timestep any of the actions can be used in any combination.

Intrinsic Reward Modules

Module	Remark	Repr.	Visual	Reference
PseudoCounts	Count-Based exploration	✔️	✔️	Never Give Up: Learning Directed Exploration Strategies
ICM	Curiosity-driven exploration	✔️	✔️	Curiosity-Driven Exploration by Self-Supervised Prediction
RND	Count-based exploration	❌	✔️	Exploration by Random Network Distillation
GIRM	Curiosity-driven exploration	✔️	✔️	Intrinsic Reward Driven Imitation Learning via Generative Model
NGU	Memory-based exploration	✔️	✔️	Never Give Up: Learning Directed Exploration Strategies
RIDE	Procedurally-generated environment	✔️	✔️	RIDE: Rewarding Impact-Driven Exploration for Procedurally-Generated Environments
RE3	Entropy Maximization	❌	✔️	State Entropy Maximization with Random Encoders for Efficient Exploration
RISE	Entropy Maximization	❌	✔️	Rényi State Entropy Maximization for Exploration Acceleration in Reinforcement Learning
REVD	Divergence Maximization	❌	✔️	Rewarding Episodic Visitation Discrepancy for Exploration in Reinforcement Learning

• 🐌: Developing.
• Repr.: The method involves representation learning.
• Visual: The method works well in visual RL.

See Tutorials: Use Intrinsic Reward and Observation Augmentation for usage examples.

Benchmark

rllte provides a large number of reusable bechmarks, see https://hub.rllte.dev/ and https://docs.rllte.dev/benchmarks/

API Documentation

View our well-designed documentation: https://docs.rllte.dev/

How To Contribute

Welcome to contribute to this project! Before you begin writing code, please read CONTRIBUTING.md for guide first.

Acknowledgment

This project is supported by The Hong Kong Polytechnic University, Eastern Institute for Advanced Study, and FLW-Foundation. EIAS HPC provides a GPU computing platform, and Ascend Community provides an NPU computing platform for our testing. Some code of this project is borrowed or inspired by several excellent projects, and we highly appreciate them. See ACKNOWLEDGMENT.md.