rl-practice-parl

Reinforcement learning practices with framework PARL.

PARL is a flexible and high-efficient reinforcement learning framework.

This repository is inspired by the one week open course of PARL. It's a wonderful course for beginners of reinforcement learning and those who want a good RL framework for practice or research.

As a newbie of both reinforcement learning and PARL myself, I followed the course lives at night after work. During the week of course, I tried Sarsa, Q-learning, DQN, Policy Gradient and DDPG in OpenAI Gym and RLSchool environments. After the course ended, I dived into a few more cases during weekends for the "Final Reproduce Tasks".

In this repository, I put all the codes of my homework and the new cases I tried with PARL. You can see it's really simple to run a RL project with PARL, only a few lines of codes modifications are needed for running different projects. Also, as reinforcement learning takes a long time to train, some pre-trained models are presented and you can download and see how it works in a minute!

codes4courses

It contains the notebooks for homework projects. For projects #4 and #5, pre-trained models are included.

• 1. gridworld
• 2. maze
  • 2.1 maze-sarsa
  • 2.2 maze-q-learning
• 3. mountaincar-dqn
• 4. pong-pg
• 5. quadrotor-hovering-ddpg

RL-Quadrotor

Notebook, codes and results of Quadrotor Velocity Control task in RLSchool.

• "velocity_control" task

Yellow arrow is the expected velocity vector; orange arrow is the real velocity vector.

RL-FlappyBird

Codes and results of Flappy Bird in PLE pygame.

• FlappyBird

RL-BipedalWalker

Notebook, codes and results of BipedalWalker task in openAI gym.

• BipedalWalker

Installation

First install requirements:

pip install -r requirements.txt

Then go to RL-Quadrotor or RL-FlappyBird or RL-BipedalWalker and try with:

python train.py