Fede wrote:
How large is the error on the control group?
First, a word on the error:
For one training entry, the output should be (1,0) for game played by player A, or (0,1) for games played by player B. And I defined the error simply as the average error for each output. So if the network output (x,y) instead of (1,0), the error is (abs(1-x)+abs(0-y))/2
Then, I defined the error on the training group or control group simply as the average error for all training entries.
Usually, I would use the quadratic error for each training entry. But since one can expect to have a lot of duplicate entries for both player A and player B (all the games they start by hoshi in top corner...) I don't want to give to much weight than necessary to those entries. So simple average.
So in my last run, the training group reached 3% average error, and the control group was at 43%
Fede wrote:
What is the bottleneck at the time being? The games played by A or the games for which we need to establish whether the player was A or B?
If it's the former, I'm still searching new games to use to train the network for A, so the number of games available could increase. It would be nice to have an idea on how many would be needed.
I am not sure in fact. I guess more training data will make it harder for the network to over-fit. But it might just make it impossible to converge as well.
I started a run with 2x400 training games and 2x400 control games. It's very slow on my computer, so I will let it run this night. If it over-fits, then I don't know what to do. If it does not converge, I will try with bigger network sizes (more layers).
If it converges but don't over-fit, bingo
In any case, I will share the training data (or the way to generate them from SGF) to let specialists have a try.
_________________
I am the author of
GoReviewPartner, a small software aimed at assisting reviewing a game of Go. Give it a try!