katago 1.4.2 (top open-source 9x9 and 19x19 bot?)

--===============1931967607224723463==
Content-Type: multipart/alternative; boundary="00000000000019be0805a5c47421"

--00000000000019be0805a5c47421
Content-Type: text/plain; charset="UTF-8"

@Hiroshi Yamashita - about rn - cool! Really neat to see people picking up
some of KataGo's methods and running them. Hopefully people will find ways
to improve them further.

------

Also, in case people weren't aware since I hadn't advertised it on this
list, KataGo has also been available for a while too! And has recently just
had its 1.4.2 release!
https://github.com/lightvector/KataGo/releases

At this moment, KataGo may be the overall-strongest open source bot on all
of 9x9, 13x13, and 19x19. Or I hope if not literally the strongest, then
very closely in the running.

* For 9x9, the version that topped
http://www.yss-aya.com/cgos/9x9/standings.html not too long ago is simply
just KataGo's v1.3.5 release (now superseded by v1.4.2) and using the 40
block neural net from here:
https://github.com/lightvector/KataGo/releases/tag/v1.4.0.

* For 19x19, the same neural net plays at a very high level too, generally
stronger than Leela Zero (except for mi yuting's flying dagger). It varies
by hardware and different time settings and configurations, but somewhere
from 100-250 Elo would probably be typical to observe in a given test[1].

* KataGo also plays all intermediate sizes - 13, 15, even stuff like 12 -
at just as high of a level. Or it should, since it trains on them all the
same way. But it doesn't seem like there's notable competition going on on
those sizes, even on 13x13.

[1]Funnily enough, the best 19x19 tests have been with KataGo's 20 block
network, *not* the 40 block network - the 40 block network is stronger
per-playout, but the 20 block (which has been learning from the 40 block's
games) is faster by enough to more than make up for it. Although actually
the 40 block network has recently made large gains and might have caught up
at time parity on 19x19 finally. There hasn't been enough testing to know
for sure yet.

On Sat, May 16, 2020 at 9:43 AM Hiroshi Yamashita <[email protected]> wrote:

> Hi,
>
> Kensuke Matsuzaki released rn-6.3.0.
> It is one of the strongest 9x9 engine.
> It is "LeelaZero + 9x9 + heuristic features + adjustable komi + KataGo
> like learning".
>
> rn-6.3.0
> https://github.com/zakki/leela-zero/releases/tag/rn-6.3.0
> Rn.6.3
> https://twitter.com/k_matsuzaki/status/1260908554359173120
>
> Author says v995 is a latest model, but v945 is stronger.
> And v995 tends to play (4,4) on initial position.
>
> Thanks,
> Hiroshi Yamashita
>
>
> This is quote from README.md in zip file.
> -----------------------------------------------------------------------
> # '9x9-endstate' branch
>
> * For 9x9 game.
> * Ladder detection (by https://github.com/yssaya/leela-zero-ladder)
> * Various komi (by https://github.com/ihavnoid/leela-zero)
> * Additional input features.
>
> -----------------------------------------------------------------------
>
> # 'Endstate' branch
>
> This is a fork of Leela Zero with the 'endstate' head.  The 'stock' Leela
> Zero uses the value and policy nets, while this also
> predicts how the game ends.  To do so, there are some changes:
>
> * Additional 'endstate' head : The 'endstate' is how the game ended - that
> information is also stored on the training data.
> * Acceleration mode : To predict the endstate, we can't just resign when
> we find the game hopeless - we have to play it to the end.  Hence,
>    once we hit the resignation threshold, we reduce the playouts to 1
> instead of resigning.
> * Using the 'endstate' information as the auxillary policy - see
> Network.cpp for details on how it uses the auxilary policy
>
> The main goal of this branch is to play reasonable handicap games (and to
> some extent, play games with komi)
> -----------------------------------------------------------------------
> _______________________________________________
> Computer-go mailing list
> [email protected]
> http://computer-go.org/mailman/listinfo/computer-go
>

--00000000000019be0805a5c47421
Content-Type: text/html; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

<div dir=3D"ltr"><div dir=3D"ltr">@Hiroshi Yamashita - about rn - cool! Rea=
lly neat to see people picking up some of KataGo&#39;s methods and running =
them. Hopefully people will find ways to improve them further.<div><br></di=
v><div>------</div><div><br></div><div>Also, in case people weren&#39;t awa=
re since I hadn&#39;t advertised it on this list, KataGo has also been avai=
lable for a while too! And has recently just had its 1.4.2 release!=C2=A0</=
div><div><a href=3D"https://github.com/lightvector/KataGo/releases">https:/=
/github.com/lightvector/KataGo/releases</a>=C2=A0=C2=A0<br></div><div><br><=
/div><div>At this moment, KataGo may be the overall-strongest open source b=
ot on all of 9x9, 13x13, and 19x19. Or I hope if not literally the stronges=
t, then very closely in the running.</div><div><br></div><div>* For 9x9, th=
e version that topped=C2=A0<a href=3D"http://www.yss-aya.com/cgos/9x9/stand=
ings.html">http://www.yss-aya.com/cgos/9x9/standings.html</a>=C2=A0not too =
long ago is simply just KataGo&#39;s v1.3.5 release (now superseded by v1.4=
.2) and using the 40 block neural net from here:=C2=A0<a href=3D"https://gi=
thub.com/lightvector/KataGo/releases/tag/v1.4.0">https://github.com/lightve=
ctor/KataGo/releases/tag/v1.4.0</a>.=C2=A0</div><div><br></div><div>* For 1=
9x19, the same neural net plays at a very high level too, generally stronge=
r than Leela Zero (except for mi yuting&#39;s=C2=A0flying dagger). It varie=
s by hardware and different time settings and configurations, but somewhere=
 from 100-250 Elo would probably be typical to observe in a given test[1].<=
/div><div><br></div><div>* KataGo also plays all intermediate sizes - 13, 1=
5, even stuff like 12 - at just as high of a level. Or it should, since it =
trains on them all the same way. But it doesn&#39;t seem like there&#39;s n=
otable competition going on on those sizes, even on 13x13.</div><div><br></=
div><div><font size=3D"1">[1]Funnily enough, the best 19x19 tests have been=
 with KataGo&#39;s 20 block network, <i>not</i>=C2=A0the 40 block network -=
 the 40 block network is stronger per-playout, but the 20 block (which has =
been learning from the=C2=A040 block&#39;s games) is faster by enough to mo=
re than make up for it. Although actually the 40 block network has recently=
 made large gains and might have caught up at time parity on 19x19 finally.=
 There hasn&#39;t been enough testing to know for sure yet.</font></div><di=
v><br></div></div><div class=3D"gmail_quote"><div dir=3D"ltr" class=3D"gmai=
l_attr">On Sat, May 16, 2020 at 9:43 AM Hiroshi Yamashita &lt;<a href=3D"ma=
ilto:[email protected]">[email protected]</a>&gt; wrote:<br></div><blockquote=
 class=3D"gmail_quote" style=3D"margin:0px 0px 0px 0.8ex;border-left:1px so=
lid rgb(204,204,204);padding-left:1ex">Hi,<br>
<br>
Kensuke Matsuzaki released rn-6.3.0.<br>
It is one of the strongest 9x9 engine.<br>
It is &quot;LeelaZero + 9x9 + heuristic features + adjustable komi + KataGo=
 like learning&quot;.<br>
<br>
rn-6.3.0<br>
<a href=3D"https://github.com/zakki/leela-zero/releases/tag/rn-6.3.0" rel=
=3D"noreferrer" target=3D"_blank">https://github.com/zakki/leela-zero/relea=
ses/tag/rn-6.3.0</a><br>
Rn.6.3<br>
<a href=3D"https://twitter.com/k_matsuzaki/status/1260908554359173120" rel=
=3D"noreferrer" target=3D"_blank">https://twitter.com/k_matsuzaki/status/12=
60908554359173120</a><br>
<br>
Author says v995 is a latest model, but v945 is stronger.<br>
And v995 tends to play (4,4) on initial position.<br>
<br>
Thanks,<br>
Hiroshi Yamashita<br>
<br>
<br>
This is quote from README.md in zip file.<br>
-----------------------------------------------------------------------<br>
# &#39;9x9-endstate&#39; branch<br>
<br>
* For 9x9 game.<br>
* Ladder detection (by <a href=3D"https://github.com/yssaya/leela-zero-ladd=
er" rel=3D"noreferrer" target=3D"_blank">https://github.com/yssaya/leela-ze=
ro-ladder</a>)<br>
* Various komi (by <a href=3D"https://github.com/ihavnoid/leela-zero" rel=
=3D"noreferrer" target=3D"_blank">https://github.com/ihavnoid/leela-zero</a=
>)<br>
* Additional input features.<br>
<br>
-----------------------------------------------------------------------<br>
<br>
# &#39;Endstate&#39; branch<br>
<br>
This is a fork of Leela Zero with the &#39;endstate&#39; head.=C2=A0 The &#=
39;stock&#39; Leela Zero uses the value and policy nets, while this also<br=
>
predicts how the game ends.=C2=A0 To do so, there are some changes:<br>
<br>
* Additional &#39;endstate&#39; head : The &#39;endstate&#39; is how the ga=
me ended - that information is also stored on the training data.<br>
* Acceleration mode : To predict the endstate, we can&#39;t just resign whe=
n we find the game hopeless - we have to play it to the end.=C2=A0 Hence,<b=
r>
=C2=A0 =C2=A0once we hit the resignation threshold, we reduce the playouts =
to 1 instead of resigning.<br>
* Using the &#39;endstate&#39; information as the auxillary policy - see Ne=
twork.cpp for details on how it uses the auxilary policy<br>
<br>
The main goal of this branch is to play reasonable handicap games (and to s=
ome extent, play games with komi)<br>
-----------------------------------------------------------------------<br>
_______________________________________________<br>
Computer-go mailing list<br>
<a href=3D"mailto:[email protected]" target=3D"_blank">Computer-g=
[email protected]</a><br>
<a href=3D"http://computer-go.org/mailman/listinfo/computer-go" rel=3D"nore=
ferrer" target=3D"_blank">http://computer-go.org/mailman/listinfo/computer-=
go</a><br>
</blockquote></div></div>

--00000000000019be0805a5c47421--

--===============1931967607224723463==
Content-Type: text/plain; charset="us-ascii"
MIME-Version: 1.0
Content-Transfer-Encoding: 7bit
Content-Disposition: inline

_______________________________________________
Computer-go mailing list
[email protected]
http://computer-go.org/mailman/listinfo/computer-go

--===============1931967607224723463==--