Result tends to be 1 after 1 million step training

yanpanlau / Keras-FlappyBird

Using Keras and Deep Q-Network to Play FlappyBird

435 stars 194 forks source link

Result tends to be 1 after 1 million step training #21

Open XiaoZzai opened 6 years ago

XiaoZzai commented 6 years ago

Though trained weights is available in this repository, I still want to train the model from scratch by myself. But after 1 million step training, all the output is 1. And I didn't modify any source code. So I am confused . What 's wrong with me ? Or are there any tips I should use ?

Thanks for answering my questions !

yanpanlau commented 6 years ago

I just re-test it. I don't see any problem. I also update the code to work for Keras 2.x and tensorflow 1.5.0. Can you try again?

XiaoZzai commented 6 years ago

Thanks a lot and I will have a try! I also wonder have you ever try to play torcs with ddpg using image as only input. I tried but failed ...