I want to make DQN with pytorch to work with simple game like PongNoFrameskip-v4, I coded up the deep Q learning code here: https://github.com/DontGiveUpEasily/pytorch-dqn however, it runs pretty well on one machine and fails on every other machines. The machine that works pretty well has pytorch version 0.12.0 and python2.7.6 (not anaconda), while other machines have pytorch 0.2.0 and python 3.5. I downgraded pytorch version to 0.12.0 and python to 2.7 (no longer using anaconda) and tried again but it didn’t help. What could be the reason?
Did you use the same seed (both openAI gym and Pytorch) when you tried to replicate your experiments?
yes, exactly the same code, completely different results. Repeat for multiple times.