Since the random games in my first iteration still only play legal moves
the network will learn to play uniformly legal moves, and will assign 0
probability for play moves on stones.
This happens after some hours of training for 9x9 and two days of
training for 19x19 in my case. Basically the loss is very high and flat
and then it quickly drops and stay still on the new level. 19x19 is
still dropping. Iteration 2 for 9x9 has started so it play MCTS moves.
It seems that learning started right away since the loss seems to drop
slightly, but still larger than the usual random noise.
I know from previous similar self play experiments that the neural
network will learn, the questions is more to what level it will
converge. I think I had the problem that the self play procedure got
stuck on repeating the same moves over and over with very low learning
and then the network will generalize worse. I stopped that experiment
when I realized that my network based on 7.0 komi CGOS 9x9 games would
be much stronger playing 5.5 komi games than the self play based network
trained on 5.5 games.
What worries me is that the quality of my selfplay games are too low,
perhaps i need to use 10x or 100x times the time per move. A difference
is that I am using the MCTS evaluation of Odin so it will always be able
to play slightly better than the raw prior from the network.
Magnus Persson
On 2017-11-07 20:19, Xavier Combelle wrote:
I wonder if some of zero project (project based on alphago zero paper)
that if I understood well was launched
did already had gather some kind of mesurable succeed, even very only
of
the order of hundreds points.
If I understand correctly, the previous mails, the computation power
you
have is 1700 less than the one of google.
So, as alphago zero goes approx 1000 elo points on 5 hours (just a
rought mesurement on nature paper graph), one could have 2.8 ELO points
won by day.
Xavier Combelle
_______________________________________________
Computer-go mailing list
[email protected]
http://computer-go.org/mailman/listinfo/computer-go
_______________________________________________
Computer-go mailing list
[email protected]
http://computer-go.org/mailman/listinfo/computer-go
