A CNN that starts with a board and returns a single number will typically have a few fully-connected layers at the end. You could make the komi an extra input in the first one of those layers, or perhaps in each of them.
Álvaro. On Sun, Jan 11, 2015 at 10:59 AM, Detlef Schmicker <d...@physik.de> wrote: > Hi, > > I am planing to play around a little with CNN for learning who is leading > in a board position. > > What would you suggest to represent the komi? > > I would try an additional layer with every point having the value of komi. > > Any better suggestions:) > > > By the way: > Todays bot tournament nicego19n (oakfoam) played with a CNN for move > prediction. > It was mixed into the original gamma with some quickly optimized parameter > leading to >100ELO improvement for selfplay with 2000 playouts/move. I used > the Clark and Storkey Network, but with no additional features (only a > black and a white layer). I trained it on 60000 kgs games and reached about > 41% prediction rate. I have no delayed evaluation, as I evaluate no > mini-batch but only one position taking about 1.6ms on the GTX-970. A > little delay might happen anyway, as only one evaluation is done at once > and other threads might go on playing while one thread is doing CNN. We > have quite slow playouts anyway, so I had around 70000 playouts/move during > the game. > > If you want to get an impression, how such a bot plays, have a look at the > games :) > > Detlef > _______________________________________________ > Computer-go mailing list > Computer-go@computer-go.org > http://computer-go.org/mailman/listinfo/computer-go
_______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
