Page 4 of 4 FirstFirst ... 234
Results 31 to 38 of 38
Like Tree9Likes

Thread: Machine learning TD Bot

  1. #31
    Senior Member
    TheRigger is offline

    Posts
    634
    Quote Originally Posted by pas View Post
    How active is TD in puzzle pirates? Might come back and write a TD bot (using either alpha beta pruning or a machine learning / genetic hybrid or even use alpha beta to train against).
    I just need to play more TD if I do since I don't even know how the game functions properly (edge cases).
    I wrote an Alpha-Beta pruning bot that did some ungodly shit but somehow worked (it got Ult, beat the public TD bot). It was pretty good for testing Jack's bot wasn't derpy as heck, but his bot crushes mine now. AlphaGo Zero TD is OP as bananas ;-;

  2. #32
    pas
    Member
    pas is offline

    Posts
    63
    Quote Originally Posted by TheRigger View Post
    I wrote an Alpha-Beta pruning bot that did some ungodly shit but somehow worked (it got Ult, beat the public TD bot). It was pretty good for testing Jack's bot wasn't derpy as heck, but his bot crushes mine now. AlphaGo Zero TD is OP as bananas ;-;
    That's quite suprising. How deep does your TD bot look ahead?
    I knew about AlphaGo Lee version which wasn't made simply by training against itself but now I read about AlphaGo Zero and it's quite remarkable what it does without any data.

    Edit:
    more research and I see AlphaZeros even been used in chess to beat the most notiorious chess engines with only 4 hours of training! This is actually amazing
    Last edited by pas; 03-11-2018 at 10:29 PM.

  3. #33
    SF Pleb
    jak8222 is offline

    Posts
    141
    The bot can hit around 60 depth depending on the player it is playing against! Riggers hits 15 depth fairly comfortably in turbo time control

    Playing against HSF really tested it, was hitting around 71 to try and find a win against him!
    Last edited by jak8222; 03-11-2018 at 10:38 PM.

  4. #34
    Senior Member
    TheRigger is offline

    Posts
    634
    The really cool thing about Jack's bot versus my bot is that his uses MCTS while mine uses Alpha-Beta yet can still outplay top players. My 14-16 depth search allows the bot to be nitpicky (extending up to about 30-ish depth on interesting positions) about an opponent's moves. In this case this means the bot capitalizes on the mistakes made by an opponent. However, Jack's high depth search is very narrowly focused, trying to hit solutions way further in the game. This means that his bot tries to outmaneuver the opponent, which is way closer to the way a real TD player plays. In addition, it makes few to no mistakes, meaning that my bot has a very difficult time exploiting opportunities, because often they turn out to be disadvantageous or neutral later on.
    tl;dr: Jack's bot is smart

    Edit: With regards to edge cases as you (pas) mentioned earlier on, TD is chock full of them. Jack tried writing and fixing his simulator a few times and never got it right
    jak8222 likes this.

  5. #35
    Senior Member
    erik is offline

    Posts
    342
    update: no permission from Google to train on their TPUs.

    it may be several years now before Annette can take over the world.

  6. #36
    Captain Medals
    Mehfailz x3 is offline

    Posts
    1,919
    Quote Originally Posted by erik View Post
    update: no permission from Google to train on their TPUs.

    it may be several years now before Annette can take over the world.
    Can't you just rent another online cloud for a few hours? I've linked jake to a few in the past. Shouldn't take more than a few hours to get a reasonable progress.
    Who do you think you are?


  7. #37
    SF Pleb
    jak8222 is offline

    Posts
    141
    TPU's are oped as hell ~20x the speed of a regular gpu, however would need some recoding. I could keep training it but it doesn't seem worth it.

  8. #38
    Junior Member
    jmal2005 is offline

    Posts
    6
    Quote Originally Posted by jak8222 View Post
    TPU's are oped as hell ~20x the speed of a regular gpu, however would need some recoding. I could keep training it but it doesn't seem worth it.
    The source code is not linked anymore. I'm self studying machine learning and would love to take a look at the work you've done so far. I like to re-code and read through other peoples work to get a better understand of how to approach problems. Cheers

Page 4 of 4 FirstFirst ... 234

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •