Deep-Q-Network-Agents

Final project for Deep Learning independent study at CUNY Lehman. A few DQN Agents that beat different types of games.

Discrete Mountain Car Agent

• beats classic control game Mountain Car with a discrete action space
  

Acrobot Agent

• beats classic control game Acrobot with a discrete action space
  

CartPole Agent

• beats classic control game CartPole with a discrete action space.
  

Pendulum Agent

• Pendulum agent created with ActionDiscretizeWrapper() wrapper to make the continuous action space discrete allowing the DQN agent to work
  

Space Invaders Agent

• Space Invaders agent created using a ROM from atarimania.com then defining the QNetwork based on the observation space, the action space, and taking into account the frame skipping inherent to old Atari games
  

Info

• The DNQ Agents can only train on environments with a discreet action space, not a continuous action space, so continuous action spaces must be converted.
  
• Action space is the range of moves you can make as the player.
  
• Observation space is everything the agent will take into account when choosing its next action (environment).