asi has been achieved internally

q learning optimizations that worked main theme: compress states into smaller, info-dense chunks. maximize (info^2 / state space) - gave it relative distances to objects from the head, rather than absolute positions of objects - gave it log(distance) to things instead of the distance other: - scaled rewards appropriately - made learning rate smaller - having a 1% chance to randomly move has high p(killing the snake) if its running along a border or next to itself

very nice. now we need another predator

snake with two snakes would be interesting

Hey, try it with PufferLib. I bet you get big snake fast

PufferLib seems cool, but my experiments are pretty small and just for learning RL id probably look into it more if I were doing RL for a company but right now its just learning on the side

That's insane!

was thinking about doing this same project, this is dope!

it was fun figuring out optimizations. it barely got above len=2 for the longest time.

I want to see q learning on an actual complex game like league of legends or world of Warcraft.

id be interested to see how someone would pull that off successfully youd have to do a TON of compression of the state space, considering how much there would be to take into account