/RLeXplore

RLeXplore provides stable baselines of exploration methods in reinforcement learning, such as intrinsic curiosity module (ICM), random network distillation (RND) and rewarding impact-driven exploration (RIDE).

Primary LanguageJupyter NotebookMIT LicenseMIT



RLeXplore: Accelerating Research in Intrinsically-Motivated Reinforcement Learning

RLeXplore is a unified, highly-modularized and plug-and-play toolkit that currently provides high-quality and reliable implementations of eight representative intrinsic reward algorithms. It used to be challenging to compare intrinsic reward algorithms due to various confounding factors, including distinct implementations, optimization strategies, and evaluation methodologies. Therefore, RLeXplore is designed to provide unified and standardized procedures for constructing, computing, and optimizing intrinsic reward modules.

The workflow of RLeXplore is illustrated as follows:

Table of Contents

Installation

  • with pip recommended

Open a terminal and install rllte with pip:

conda create -n rllte python=3.8
pip install rllte-core
  • with git

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

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

Now you can invoke the intrinsic reward module by:

from rllte.xplore.reward import ICM, RIDE, ...

Module List

Type Modules
Count-based PseudoCounts, RND, E3B
Curiosity-driven ICM, Disagreement, RIDE
Memory-based NGU
Information theory-based RE3

Tutorials

Click the following links to get the code notebook:

  1. Quick Start
  2. RLeXplore with RLLTE
  3. RLeXplore with Stable-Baselines3
  4. RLeXplore with CleanRL
  5. Exploring Hybrid Intrinsic Rewards
  6. Custom Intrinsic Rewards

Benchmark Results

We have published a space using Weights & Biases (W&B) to store reusable experiment results on recognized benchmarks. The space link is: RLeXplore's W&B Space.

  • RLLTE's PPO+RLeXplore on SuperMarioBros:

  • RLLTE's PPO+RLeXplore on MiniGrid:

    • DoorKey-16×16
    • KeyCorridorS8R5, KeyCorridorS9R6, KeyCorridorS10R7, MultiRoom-N7-S8, MultiRoom-N10-S10, MultiRoom-N12-S10, Dynamic-Obstacles-16x16, and LockedRoom

  • RLLTE's PPO+RLeXplore on Procgen-Maze:

    • Number of levels=1
    • Number of levels=200

  • RLLTE's PPO+RLeXplore on five hard-exploration tasks of ALE:

Algorithm Gravitar MontezumaRevenge PrivateEye Seaquest Venture
Extrinsic 1060.19 42.83 88.37 942.37 391.73
Disagreement 689.12 0.00 33.23 6577.03 468.43
E3B 503.43 0.50 66.23 8690.65 0.80
ICM 194.71 31.14 -27.50 2626.13 0.54
PseudoCounts 295.49 0.00 1076.74 668.96 1.03
RE3 130.00 2.68 312.72 864.60 0.06
RIDE 452.53 0.00 -1.40 1024.39 404.81
RND 835.57 160.22 45.85 5989.06 544.73
  • CleanRL's PPO+RLeXplore's RND on Montezuma's Revenge:
  • RLLTE's SAC+RLeXplore on Ant-UMaze:

Cite Us

To cite this repository in publications:

@article{yuan_roger2024rlexplore,
  title={RLeXplore: Accelerating Research in Intrinsically-Motivated Reinforcement Learning},
  author={Yuan, Mingqi and Castanyer, Roger Creus and Li, Bo and Jin, Xin and Berseth, Glen and Zeng, Wenjun},
  journal={arXiv preprint arXiv:2405.19548},
  year={2024}
}