I am attempting to solve an OpenAI Gym environment using Deep Q Learning with TensorFlow on an M1 Mac. The model seems to start fitting well but ends up performing worse every time. In contrast, when I used the same code in Google Colab, it worked successfully. Could someone please explain what the potential problem might be? Could it be related to the different types of TensorFlow, given that the M1 Mac uses TensorFlow Metal and TensorFlow macOS?
I ran the same code on google colab using same gym version and tensorflow(default version). The model started training as it was on my mac, but it trained as I expected whereas on my m1 mac, the scores of the agent (for lunar lander) starts from -180ish or even -200 which is totally fine, it then starts to go down(going positive) every episodes(same on colab). But after a few hundred episodes the score starts going back in negative direction(colab model goes positive till it reaches +200 and the model is saved).
