search

harbecke / HexHex

AlphaGo Zero adaptation for Hex
GNU General Public License v3.0
20 stars 4 forks source link

data is in overcomplicated format #15

Closed harbecke closed 5 years ago

harbecke commented 5 years ago

board_tensor has three channels, where one or two should suffice. the tensor should be transposed after a player made a move, such that the model only plays the board in one direction. This removes the need for VerticalWrapperModel

PascalCremer commented 5 years ago

I support the change but I would be very interested to see if the VerticalWrapperModel actually improves the performance on e.g. 7x7.

harbecke commented 5 years ago

do you have a specific test in mind or should i just train a model for 300 epochs without the VerticalWrapperModel and we compare?

cleeff commented 5 years ago

That is exactly what I had in mind. Would be great to have some of the comparison models in reference_models.json

harbecke commented 5 years ago

@cleeff @simonant suggestion:

  1. two channels in the current format (channel 0=vertical player, channel 1=horizontal player)
  2. combine these channels into one channel (1 indicates stone of vertical player, -1 of horizontal player)

in the first version the tensor gets transposed and the channels are switched after a move, in the second version it gets transposed and multiplied with (-1)

i'll test which of these versions works better. or do you see pitfalls?

PascalCremer commented 5 years ago

Sounds sane.

harbecke commented 5 years ago
  1. is implemented with 58a460ebb3963201b5632e72fb59f6076d59491f