Re: [Computer-go] Patterns and bad shape

[lemonsqueeze]

I think that's ok: the prediction systems are already used to deal with a huge number of positions during training, it's just a matter of changing the quality of these positions. Say instead of training on 100% good answers to good moves from games, we could take half as many and train on 50% positions from games + 50% good answers to bad moves. Some systems scale really well so we could also just double the training set. Even if it hurts prediction rate a little bit, having good answers to bad moves seems crucial for a good search.

      A B C D E F G H J K L M N O P Q R S T
    +---------------------------------------+
 19 | . . . . . . . . . . . . . . . . . . . |
 18 | . O . . . . . . . . . . . . . . . . . |
 17 | . X O O . O . . . . . . . X . X . . . |
 16 | . X X O . . . . . . . . . . . . X . . |
 15 | . . . X O . . . . . . . . . . O X . . |
 14 | . . X . . . . . . . . . . . . O O . . |
 13 | . . O O . . . . . . . . . . . . . . . |
 12 | . . O X X . . . . . . . . . . . . . . |
 11 | . . O O X . . . . . . . . . . . . . . |
 10 | . . X X X O . . . . . . . . . . O . . |
  9 | . . . . X . . . . . . . . . . . . . . |
  8 | . . O . . . . . . . . . . . . . X . . |
  7 | . . . . . . . . . . . . . . . . . . . |
  6 | . . . . . . . . . . . . . . . . . . . |
  5 | . . O . . . . . . O . . . . . . . . . |
  4 | . . . . . . . X)O X X . . . . . X . . |
  3 | . . . O O . . X X O O X . . X . . . . |
  2 | . . . X . . . O O O O X . . . . . . . |
  1 | . . . . . . . . . O X X . . . . . . . |
    +---------------------------------------+



  [ J5   M4   G3   G2   H5   ...
  [ 0.92 0.05 0.03 0.01 0.00 ...

Matt


On 04/17/2017 01:04 PM, Jim O'Flaherty wrote:

Btw, I thought neural networks trained from games were ok with basic bad shapes because they "understand" the situation much better, but in some cases they still make incredible local blunders, so I guess they're affected too. Here's an example with Detlef's 54% dcnn: Network is almost 100% sure that saving 1 stone in atari with J5 is the right move here, and loses all bottom 7 stones !

It seems chasing down good moves for bad shapes would be an explosion of
"exception cases", like combinatorially huge. So, while you would be
saving some branching in the search space, you would be ballooning up
the number of patterns for which to scan by orders of magnitude.

Wouldn't it be preferable to just have the AI continue to make the
better move emergently and generally from probabilities around win
placements as opposed to what appears to be a focus on one type of local
optima?

game.sgf
Description: application/go-sgf

_______________________________________________
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go