1. This site uses cookies. By continuing to use this site, you are agreeing to our use of cookies. Learn More.

News Google build first champion Go AI

Discussion in 'Article Discussion' started by Gareth Halfacree, 28 Jan 2016.

  1. Gareth Halfacree

    Gareth Halfacree WIIGII! Staff Administrator Super Moderator Moderator

    Joined:
    4 Dec 2007
    Posts:
    11,561
    Likes Received:
    1,367
  2. theshadow2001

    theshadow2001 [DELETE] means [DELETE]

    Joined:
    3 May 2012
    Posts:
    5,147
    Likes Received:
    136
    So are they going to use this AI to enable their customers to sell more crap to internet users or do they have an actual use for it?
     
  3. Gareth Halfacree

    Gareth Halfacree WIIGII! Staff Administrator Super Moderator Moderator

    Joined:
    4 Dec 2007
    Posts:
    11,561
    Likes Received:
    1,367
    Yes, the use is "convincing people to use Google DeepMind over rival deep-learning platforms." See also: the main reason Mercedes (or A. N. Other car maker) has a Formula One team.

    You've got to admit, it's a pretty good advert for DeepMind. Look at the details: they didn't teach it how to play Go, they just gave it a copy of Go and told it to go to town, after prepping it with generic algorithms. The AI learned how to win all by itself. That's some pretty impressive stuff, and if I were doing soft-AI stuff it'd make me want to have a look at using DeepMind.
     
  4. Cthippo

    Cthippo Can't mod my way out of a paper bag

    Joined:
    7 Aug 2005
    Posts:
    6,783
    Likes Received:
    101
    Now if someone would just apply this to dating site algorythms, we could be in business.
     
  5. Ending Credits

    Ending Credits Bunned

    Joined:
    4 Jan 2008
    Posts:
    5,234
    Likes Received:
    210
    Gotta correct you there, no Genetic Algorithms are involved in the training of AlphaGo.

    It's initially taught how to play by Supervised Learning on a bunch of previous human games (so it can predict moves a human might make in a given situation). Then this network is trained using an unorthodox form of reinforcement learning by playing games against previous versions against itself and updating network weights depending on whether it won or lost (which is a gradient method rather than a genetic algorithm method).

    At this point it's already pretty good, but they then take the final 'policy' network and use it to train a 'value' network which evaluates the strength of various board positions. Finally this is combined with a tree search (with values of the 'value' network used to approximate weights of truncated branches) similar to most Chess engines.

    If you havent seen some of their previous work on playing a bunch of Atari games, you should check those out too. Papers are all available on the Deepmind website.
     
  6. Gareth Halfacree

    Gareth Halfacree WIIGII! Staff Administrator Super Moderator Moderator

    Joined:
    4 Dec 2007
    Posts:
    11,561
    Likes Received:
    1,367
    Dude... Re-read that word. Hint: it doesn't say "genetic". :p
     
  7. theshadow2001

    theshadow2001 [DELETE] means [DELETE]

    Joined:
    3 May 2012
    Posts:
    5,147
    Likes Received:
    136
    Google's technology is always impressive, except Glass.
     
  8. Ending Credits

    Ending Credits Bunned

    Joined:
    4 Jan 2008
    Posts:
    5,234
    Likes Received:
    210
    Ah, very good point lol.
     

Share This Page