Atari Game Playing

Apply DQN to Atari games: convert pixel observations to feature vectors, train on massive replay buffers, achieve superhuman performance.

The problem, from first principles

Raw pixels as state: 84x84x4 = 28,224 inputs per frame. Fully connected networks fail. Convolutional networks extract features. DQN learns to play from pixels alone.

From theory to code

Implement preprocess_atari_frame(raw_frame) which:

• Takes raw 210x160x3 RGB frame
• Crops to 160x160
• Resizes to 84x84
• Converts to grayscale
• Normalizes to [0, 1]
• Returns preprocessed frame

Constraints

• Input: 210x160x3 uint8 Atari frame
• Output: 84x84 float32 frame
• Use standard preprocessing

Hints