DRL_Implementation

Current status: minimal updates

Introduction

• This repository is a pytorch-based implementation of modern DRL algorithms, designed to be reusable for as many Gym-like training environments as possible
• The package is mainly for my personal usage, however feel free to use it as you like.
• It is recommended to use the released version
• Understand more with the Wiki!
• Tested environments: Gym, Pybullet-gym, Pybullet-multigoal-gym
• My priority is on continuous action algorithms as I'm working on robotics

Installation

git clone https://github.com/IanYangChina/DRL_Implementation.git
cd DRL_Implementation
python -m pip install -r requirements.txt
python -m pip install .

Click here for example codes , to run the codes you will need to install Gym, Pybullet, or pybullet-multigoal-gym. See env installation links below. For more use cases, have a look at the drl_imp_test repo
From the project root, run python drl_implementation/examples/$SCTIPT_NAME.py

State-based

• Distributional DDPG, Continuous
• DDPG - Deterministic, Continuous
• TD3 -Deterministic, Continuous
• SAC (Adaptive Temperature) - Stochastic, Continuous

Replay buffers

• Hindsight
• Prioritised

Tested Environments

• Pybullet Gym (Continuous)
• OpenAI Gym Mujoco Robotics Multigoal Environment (Continuous)
• Pybullet Multigoal Gym (OpenAI Robotics Multigoal Pybullet Migration) (Continuous)

Some result figures

Reference Papers: Algorithm

• DQN
• DoubleDQN
• DDPG
• TD3
• SAC (Adaptive Temperature)
• PER
• HER

Reference Papers: Implementation Matters

• Time limit
• SOTA PPO Hyperparameters (many applicable to other algorithms)
• SAC Temperature Auto-tuning