Thanks for the link! It looks like I've got another paper to put on my reading list
I've got a few papers in hand already and I think I'll make sure I read your paper before I get into actually coding some of this stuff.
On my list, I've got:
One thing I'll need to address as I get going, is that I think doing reinforcement learning with any sort of tree search component may require modifications to the game engine. I'd like to be able to submit a few candidate move assignments from a particular game state, for instance, and then run my AI in all of those.
Other obvious problems I'm sure people are thinking about, even if there are only 10 agents on the board, and you restrict your search to a discrete space of 8 compass directions at full thrust, you've got 2^30 next states to consider, which makes search pretty hairy! I'm interested in the paper I linked about tree search in continuous spaces, but I feel like search a 20 dimensional float-valued space isn't going to be easier than the discrete approximation. I'm hoping that using kernel regression helps out by enforcing a geometric notion of similarity, but I don't really know what I'm doing (but I'm excited to learn!)
Has anyone else had thoughts about the game engine thing? Is there some existing ability to do what I think I need to do that I've simply overlooked?