flappybird-DQN

This is a submission of the second training epoch of Flappy Bird using Deep Q-Learning, built with Python and Tensorflow.

The first epoch was trained with the following settings:

First buffer with 1000 random iterations
training episodes: 20000 
learning rate for Adam Optimizer: 1e-5

we used our sinusoidal epsilon function with parameters:
starting epsilon: 1.0
epsilon decay rate: .9998 
number of epsilon cycles: 7

replay memory size: 75000
batch size: 32


The second epoch was trained with the following settings:

First buffer with 5000 forward passes through the network without updates
training episodes: 20000 
learning rate for Adam Optimizer: 9e-7

we used our sinusoidal epsilon function with parameters:
starting epsilon: .2
epsilon decay rate: .9997
number of epsilon cycles: 4

replay memory size: 75000
batch size: 32


More details about the network architecture, sinusoidal decay function and other can be found in our paper (coming soon).
Link to Poster: <http://web.stanford.edu/~chuchro3/projects/openaigym/OpenAiGymDeepLearningPoster.pdf>