I am currently working on a project to build and train a neural network to play DotA 2 (at least be able to play like a really bad human).
The thing is all the neural network I have built before were made to play small games like 2048 or snake so the inputs were pretty straight-forward. (e.g for 2048 each input neuron was a tile of the game and the value was normalized from 2-2048 to 0-1).

But for DotA2 things get really more complicated because I have to encode a huge amount of possible objects: heros, creeps, towers, etc...
At the moment I am thinking about maybe dividing a simplified version of the map into tiles (about 3000 tiles) and have each tile related to 1 input but then again, how to encode such complex data ?

Hoping to read from some of you experts out here,
Thank you

  • 1
    $\begingroup$ This is a rather broad question. I suggest finding a mentor. $\endgroup$ – Yuval Filmus Sep 22 '17 at 9:31
  • $\begingroup$ I'd like to but I don't have anyone $\endgroup$ – Nifil Sep 24 '17 at 12:47
  • $\begingroup$ OpenAI have a model for 1v1 dota using only a single hero and with limited items that took months to train. $\endgroup$ – Nicholas Mancuso Sep 25 '17 at 22:07
  • $\begingroup$ Yeah I know that but it doesn't help me figure out anything $\endgroup$ – Nifil Sep 26 '17 at 12:05

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.