Re: [Computer-go] Update Odin Zero 9x9 Komi 7.
I think might be extra interesting since I train networks for 5.5 and 7.0 komi on 9x9, I have a theory that the 7.0 komi network can become much stronger because most jigo games has to played perfectly but with a move ordering the puts maximum pressure on the opponent. With 5.5 komi most games end with resignation after playing desperate moves. A sign that this matter when trained on self play games would be that the 7.0 network beats the 5.5 network playing with 5.5 komi. Cheers Magnus On 2017-11-11 07:36, Ingo Althöfer wrote: Hi Magnus, thanks also to you for your experiments. In particular, I like you setting with integral komi. Cheers, Ingo. Gesendet: Samstag, 11. November 2017 um 00:48 Uhr Von: valky...@phmp.se An: Computer-go@computer-go.org Betreff: [Computer-go] Update Odin Zero 9x9 Komi 7. Odin has now generated 5 games with 100 simulation and 90% prior for a random move, and 1 games with 5000 simulations and 15% prior for a random move, with temperature 1. ... ___ 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
Re: [Computer-go] Update Odin Zero 9x9 Komi 7.
Hi Magnus, thanks also to you for your experiments. In particular, I like you setting with integral komi. Cheers, Ingo. > Gesendet: Samstag, 11. November 2017 um 00:48 Uhr > Von: valky...@phmp.se > An: Computer-go@computer-go.org > Betreff: [Computer-go] Update Odin Zero 9x9 Komi 7. > > Odin has now generated 5 games with 100 simulation and 90% prior for > a random move, and 1 games with 5000 simulations and 15% prior for a > random move, with temperature 1. > ... ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
[Computer-go] Update Odin Zero 9x9 Komi 7.
Odin has now generated 5 games with 100 simulation and 90% prior for a random move, and 1 games with 5000 simulations and 15% prior for a random move, with temperature 1. The network has been trained all the time and the loss function is now dropping about 0.02 units every 10 hours or so. I loaded all the self generated games in in Drago and compared the distribution of moves several opening positions and compared to the priors generated by latest save network. I was expecting chaos and randomness but it already looks really ordered. It is clear that this network is not memorizing positions because the network is predicting much stronger moves than the statistics from a given position in the training games would do. I guess this also happens when the training set is less random but in this case it is really striking. I just added add some 3 versions of Odin to CGOS 9x9 all playing with 5000 simulations, Odin_1.1.8_5K without network, Odin_1.1.8_5K_N1 with the first network, and Odin_1.1.8_5K_N22 with the latest network. Magnus Persson ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go