Neural Networks with Keras Cookbook

Getting ready

The strategy that we'll adopt to build an agent that is able to maximize the score is as follows:

• Initialize the environment of the Space Invaders-Atari2600 game.
• Preprocess the image frame:
  • Remove pixels that do not necessarily impact the action prediction
    • For example, pixels below the location of the player
  • Normalize the input image.
  • Resize the image before passing it to the neural network model
• Stack frames as required by the Gym environment
• Let the agent play the game over multiple episodes:
  • During the initial episodes, we'll have high exploration which decays over increasing episodes.
  • The action that needs to be taken in a state depends on the value of the exploration coefficient.
  • Store the game state and the corresponding ...