One easy (maybe?) tweak to the novelty reward function
apple-ihack-geek opened this issue · 3 comments
What if instead of or in addition to the new screen novelty/exploration reward, you used the AI's coordinates in the game? Since you went through all the work of stitching the map into a global coordinate system for the video, I think it would also make sense to use that as part of its training and reward data. That would easily fix the current problem where the AI get's stuck in a long uniform tunnel or path.
I would dig into it myself, but I'm about to go off grid for the weekend.
This has been implemented, and can be used to get past mt moon! you can enable it by setting use_screen_explore: False in the config.
Based on the timing, I'm guessing you had already been working on this or already completed it. Either way, glad to know I
was thinking along the same lines. Maybe you don't want to spoil it and keep it hidden for a follow up video, but how far does it get? what does it hang on now?
This has been implemented, and can be used to get past mt moon! you can enable it by setting
use_screen_explore: Falsein the config.
Do I need to stop my current session and start a new one to implement this?