Results 1 to 3 of 3

Thread: For any chess or AI (or maybe Skynet) enthusiasts...

  1. #1
    Senior Member Lanky123's Avatar
    Join Date
    Jul 2007
    Location
    Oxford
    Posts
    922
    Thanks
    91
    Thanked
    152 times in 101 posts
    • Lanky123's system
      • Motherboard:
      • Gigabyte GA-H81M-D2V
      • CPU:
      • Core i5 4570
      • Memory:
      • 2 x 4GB Vengeance LP
      • Storage:
      • 250GB Samsung 840 EVO SSD + 2+4TB HDD + 3TB Synology DS216SE
      • Graphics card(s):
      • MSI Radeon R9 270X HAWK
      • PSU:
      • Silverstone Strider 400W
      • Case:
      • Silverstone Sugo SG02B-F
      • Operating System:
      • Windows 8.1 / Ubuntu 16.04
      • Monitor(s):
      • ElectriQ 32" 4k IPS + Dell 22" U2212HM
      • Internet:
      • Virgin 60Mbit/s

    For any chess or AI (or maybe Skynet) enthusiasts...

    Google has let DeepMind AlphaZero loose with a chess board, and it pretty well destroyed the previous state-of-the-art chess engine (Stockfish 8). 100 games, 28 wins, 72 draws, 0 losses. I guess it wasn't even worth letting a human try to play it. Previous engines relied on brute force tree search algorithms with some pruning and various opening/endgame lookup tables. AlphaZero is DeepMind's deep convolutional neural network tech and was apparently trained only with the rules of chess and 24 hours of self-play, no lookup tables or anything.

    The best part for me is that AlphaZero seems to play much more aesthetic or 'human-like' chess, happy to make plenty of material sacrifices because it must have a much more robust understanding of the positional advantages it can gain. On page 6 of the paper the DeepMind team show the openings that AlphaZero started playing at various points in it's self-play development - it's interesting to see some common human openings are played but then discarded as it evidently finds some flaw in them. Hundreds of years of human chess development condensed into a few hours and then completely surpassed...

    I've linked to some analysis of one of the games from a youtube channel I watch and the arxiv paper from the DeepMind team:

    https://youtu.be/lb3_eRNoH_w

    https://arxiv.org/pdf/1712.01815.pdf
    Last edited by Lanky123; 08-12-2017 at 05:50 PM.

  2. Received thanks from:

    Jowsey (09-12-2017),MLyons (08-12-2017),MrRockliffe (12-12-2017),virtuo (12-12-2017)

  3. #2
    Senior Member
    Join Date
    Aug 2013
    Location
    North Wales
    Posts
    1,849
    Thanks
    165
    Thanked
    271 times in 202 posts
    • virtuo's system
      • Motherboard:
      • Gigabyte Aorus Master X570
      • CPU:
      • Ryzen 9 5950x
      • Memory:
      • 64Gb G.Skill TridentZ Neo 3600 CL16
      • Storage:
      • Sabrent 2TB PCIE4 NVME + NAS upon NAS upon NAS
      • Graphics card(s):
      • RTX 3090 FE
      • PSU:
      • Corsair HX850 80+ Platinum
      • Case:
      • Fractal Meshify 2 Grey
      • Operating System:
      • RedStar 3, Ubuntu, Win 10
      • Monitor(s):
      • Samsung CRG90 5140x1440 120hz
      • Internet:
      • PlusNet's best, but still poor, attempt

    Re: For any chess or AI (or maybe Skynet) enthusiasts...

    I've been dipping in to Neural Net programming in my "down" time, it's fascinating and mind-bending stuff. This one is particularly impressive given the training data is just the "rules" and nothing else.

  4. #3
    Senior Member Lanky123's Avatar
    Join Date
    Jul 2007
    Location
    Oxford
    Posts
    922
    Thanks
    91
    Thanked
    152 times in 101 posts
    • Lanky123's system
      • Motherboard:
      • Gigabyte GA-H81M-D2V
      • CPU:
      • Core i5 4570
      • Memory:
      • 2 x 4GB Vengeance LP
      • Storage:
      • 250GB Samsung 840 EVO SSD + 2+4TB HDD + 3TB Synology DS216SE
      • Graphics card(s):
      • MSI Radeon R9 270X HAWK
      • PSU:
      • Silverstone Strider 400W
      • Case:
      • Silverstone Sugo SG02B-F
      • Operating System:
      • Windows 8.1 / Ubuntu 16.04
      • Monitor(s):
      • ElectriQ 32" 4k IPS + Dell 22" U2212HM
      • Internet:
      • Virgin 60Mbit/s

    Re: For any chess or AI (or maybe Skynet) enthusiasts...

    *bump*

    There is now a free + open source equivalent of this that runs on consumer hardware - Leela Chess Zero. It seems to be inspired by the AlphaZero network but with distributed training and I think the network size is a factor of a few smaller (for now, I guess they might increase it if they hit a performance ceiling).

    It is still being trained but you can play the current version in a browser here. The 'normal' difficulty is already really good (destroys me lol) and seems to be playing very 'human like' chess as AlphaZero did. I can't get the 'hard' difficulty to work as I think the server is overloaded.

    If anyone wants to contribute to the training or download the latest version of the network to play with on a local machine you can get it from the github here. There are also instructions for getting it working with various chess GUIs. Training progress and a bunch of other info here.
    Last edited by Lanky123; 06-04-2018 at 04:12 AM.

Thread Information

Users Browsing this Thread

There are currently 1 users browsing this thread. (0 members and 1 guests)

Posting Permissions

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