Christmas status of oakfoam :) Two years deep learning in oakfoam

[Detlef Schmicker]

After carefully training a

combined policy-value net I finaly did some runs on CGOS.

http://www.yss-aya.com/cgos/19x19/bayes.html

The accounts are

NG-04 (and NG-04b with a ladder bug fixed, probably showing up two days from now)

This account does 10000 playouts/move with

param uct_expand_after 20

which means about 500 node expansions with policy-value net call.

NG-04 has an Elo of about 3100

between

                    CGOS    KGS  Botname  Time 
Aya788d_p1v0_6c12t  3235    6d   AyaMC    1min + 15sec x10
Aya786m_4c          2965    4d   AyaMC4   1min + 15sec x10

from http://computer-go.org/pipermail/computer-go/2016-June/009444.html one might think, this would be around KGS 5d?!

The other account of interest is the pure policy network account only playing the best move (thus it is deterministic, but I checked that the opponent players did not play deterministic)

NG-cnn_proe_1b with Elo 2464 which should be about 2d on kgs.

Now, that we win some games against zen it appears, that the ladder reader has to be improved :) Zen does it perfectly and we only read simple ladder ends. If before the end there are two different possibilities, which become important later, we can not read this :(

Let me know, if someone wants to work at this :)

[Steve Kroon]

Hi Detlef.

Thanks for the update, this is very pleasing performance.

I trust you had a festive Christmas!

Regards,

Steve

--
You received this message because you are subscribed to the Google Groups "Oakfoam" group.
To unsubscribe from this group and stop receiving emails from it, send an email to oakfoam+unsubscribe@googlegroups.com.
To post to this group, send email to oak...@googlegroups.com.
Visit this group at https://groups.google.com/group/oakfoam.
For more options, visit https://groups.google.com/d/optout.