dqn.flappybird

Deep Q Network for flappy bird.

Description

Based on DeepLearningFlappyBird

Changes made:

1. Feed greyscale image to model, instead of binary
2. Initial epsilon to 0.5, to converge faster
3. Add TensorBoard (tensorboard --logdir=summary), Graph
4. Visualize Q change over iterations, following idea from DeepMind's paper. Q
5. Visualize network layers. Layer

Environment:

conda.environment.yml, check this

Steps

1. train

Model are saved under saved_model. 5 models are uploaded. 2. play

1k frames are played using last trained model. Frames and states are saved under frames_states. Frames are then merged to gif (convert -delay 10 -loop 0 frames_states/frame*.png animation.gif) 3. eval

Evaluate each model on 1k states 4. visualize

Plot each layer using two ways, tensorflow and matplotlib