My attempt at DeepMindR;s baseline for playing minigames:

I’m not the first to do this, but my is a bit different in the sense that I’ve tried to implement it as cleanly and as closely to DM’s description as possible.
This includes supporting training on almost full feature list, 1×1 conv embeddings for spatial categorical features, and proper policy implementation, with all actions and their variable arguments. My implementation is also flexible as far as features/actions go, supporting configs via external json file.

I think I’m also the first to tackle the harder maps, even if only with partial success.

You can see a demo of my agent on all minigames here:
(if you’re confused why it doesn’t look like Starcraft II: I’m on Linux where they support simplified UI via Pygame).

Source link
thanks you RSS link


Please enter your comment!
Please enter your name here