This is some naive implementations of deep reinforcement learning algorithms. The purpose of this repo is to help me understand the algorithms and the code. The code is not optimized for performance. If you want to use the code for your research, please refer to the original paper and the official implementation. I verify the code with OpenAI gymnasium. The most of games that I used is LunarLander-v3, CartPole-v1 and Pendulum-v1

• Naive Implementations of Deep Reinforcement Learning
  • About
  • Table of Contents
  • Environment Preparation (torch users)
  • Run
  • Algorithms
    • Improved with Gumbel Distribution Regression from XQL
  • Reference

conda create -n rltorch pytorch torchvision torchaudio pytorch-cuda=12.1 gymnasium pyglet pygame gymnasium-box2d colorama pylint yapf tqdm 'tensorboardx>=2.5.0' 'tensorboard>2.0' pillow matplotlib scipy seaborn ipykernel -c conda-forge -c pytorch -c nvidia

This project does not provide the trained Deep Reinforcement Learning model weight.

You can start training model under conda environment by

(rltorch) > python -m <project name>.main

For example (DDPG):

(rltorch) > python -m DDPG.main

DQN	DDQN	DDPG
PPO	PPG	C51
AWR	AC	TD3
SACv1	SACv2	MPO

XAWR	XDDPG
XTD3	XSAC

• TrainMonitor and Generategif modified from coax
• https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf
• https://arxiv.org/pdf/1509.06461.pdf
• https://arxiv.org/pdf/1509.02971.pdf
• https://arxiv.org/pdf/1707.06347.pdf
• https://arxiv.org/pdf/1707.06887.pdf
• https://openreview.net/attachment?id=H1gdF34FvS&name=original_pdf
• https://proceedings.neurips.cc/paper/1999/file/6449f44a102fde848669bdd9eb6b76fa-Paper.pdf
• https://div99.github.io/XQL/