A convolutional network learns to flap by dying repeatedly

What it does This project is a Python implementation of a Deep Q-Network that learns to play Flappy Bird. It feeds stacked, preprocessed game frames into a convolutional neural network and lets the agent decide when to flap and when to do nothing. The code is largely an adaptation of existing work—specifically the FlapPyBird game engine and a prior DeepLearningVideoGames DQN implementation—tied together to target this specific bird.

The interesting bit The author tweaked the standard DQN exploration schedule: instead of starting epsilon at 1, they start at 0.1 because a high random-action rate makes the bird flap manically, hug the top of the screen, and die clumsily. They also strip the background from frames to speed up convergence, which is a nice reminder that even “general” algorithms need domain-specific babysitting.

Key highlights

• Trains on raw 80×80 grayscale pixel stacks; no hand-crafted game state required.
• Uses a three-layer convnet with max pooling and 256 fully-connected ReLU nodes.
• Epsilon-greedy exploration is tuned for Flappy Bird’s physics: high randomness causes over-flapping, so epsilon anneals from 0.1 down.
• The first 10,000 steps are purely random to seed the replay buffer before training begins.
• Explicitly built on top of FlapPyBird and DeepLearningVideoGames—more adaptation than reinvention.

Caveats

• Dependencies list TensorFlow 0.7, which is years out of date and likely a compatibility headache on modern systems.
• The README includes a manual checkpoint-path fix in the FAQ, suggesting saved model loading is brittle.
• The project is explicitly described as “highly based” on two other repos, so expect glue code rather than novel architecture.

Verdict Worth a look if you want a concise, well-documented reference implementation of DQN applied to a simple 2D game. Skip it if you are looking for a state-of-the-art agent or a dependency stack from this decade.