Reinforcement Learning

This repo is for playing with reinforcement learning algorithms. I am either using openai gym or ViZDoom as an environment.

I am trying several environments which includes:

• Mountain Car
  • Implemented simple Q-Learning
• Pacman
  • Implemented DQN with CNN architectures
• Breakout
  • Implemented DQN with experience replay and CNN architecture
  • Implemented DQN with GRU architecture
• Doom Game
  • Implemented double DQN on basic scenario ViZDoom

If you need help in setting up your system (mac only) with these environments, refer this:

Openai gym

Inorder to install openai gym, run the following command in mac terminal:

• pip install gym

Run the below command in order to use atari environments:

• pip install gym[atari]

Setup render gym- Google Colab

Inorder to render env in colab, we need to need to install some dependencies; so run following commands in colab:

• !apt-get install python-opengl -y
• !apt install xvfb -y
• !pip install pyvirtualdisplay
• !pip install piglet
• !apt-get install x11-utils

After that run the following code once:

• from pyvirtualdisplay import Display
• from IPython import display
• Display().start()

Then put the following command, just before the episode starts (need to run only once):

• img = plt.imshow(env.render('rgb_array'))

Then put the following code, where we generally put env.render():

• img.set_data(env.render('rgb_array')) # just update the data
• display.display(plt.gcf())
• display.clear_output(wait=True)

ViZDoom

Inorder to install ViZDoom, run following commands in mac terminal (1st two commands were for installing dependencies for installing ViZDoom):

• brew install cmake boost sdl2 wget
• brew cask install julia
• pip install vizdoom