yenchenlin/DeepLearningFlappyBird

Implements a convolutional neural network trained with experience replay and ε-greedy exploration, processing raw 80×80×4 grayscale frame stacks as input to output Q-values for discrete actions. The architecture uses three convolutional layers with max pooling followed by a 256-unit fully connected layer, optimized via Adam on minibatches sampled from a 500k-capacity replay buffer. Built on TensorFlow 0.7 and pygame, with custom preprocessing (background removal, frame stacking) tuned specifically for Flappy Bird's fast action cadence.

6,792 stars. No commits in the last 6 months.