search

bkkhack / hacknights

A meta-repository for finding hack night birds-of-a-feather and tracking what we learn.
https://waffle.io/bkkhack/hacknights
15 stars 4 forks source link

Self-learning Doom Bot #193

Closed tao-pr closed 6 years ago

tao-pr commented 7 years ago

Implement a reinforcement-learning algo for a classic Doom bot to find best strategies to maximise the score!

I made a repo here : https://github.com/starcolon/classic-doom-bot The simple Q-Learning algo has been implemented, up and running over Jupyter Notebook.

I'll find another more suitable algo to imeplement and test the bot. Just for fun!

aljones15 commented 7 years ago

this sounds really cool. let me know on slack if this works! would love to look at the code.

tao-pr commented 7 years ago

@aljones15 definitely!

elcolie commented 7 years ago

Then let's create bot and fight

aljones15 commented 7 years ago

@elcolie that's actually a cool idea. Doom has a lan multiplayer mode so it would be cool to see learning algorithims against learning algorithims deatth match. Might be easier in source 2 though.

tao-pr commented 7 years ago

Not so much progress made tonight :hushed: I drafted a structure for Temporal Difference learning algo but still in the middle of trying to understand how the algo works before rushing into implementation.

I also have a baseline & simpler strategy implemented based on Q-learning here in master branch. I trained it for like 600 rounds and it knows how move and shoot, but still isn't smart enough :smile:

aljones15 commented 7 years ago

Thanks for the update hope to understand your project soon. :)

aljones15 commented 7 years ago

https://keon.io/deep-q-learning/ If you didn't see this one its a Q learning in Keras tutorial.

tao-pr commented 7 years ago

@aljones15 Nice article. Deep Reinforcement Learning sounds interesting to try out at some point, perhaps after Temporal Difference. Wanna compare Deep QL against basic QL too.

elcolie commented 7 years ago

Are you ready to show?