Re: [Computer-go] Zen lost to Mi Yu Ting

2017-03-21 Thread Álvaro Begué 
I was thinking the same thing. You can easily equip the value network with
several outputs, corresponding to several settings of komi, then train as
usual.

The issue with Japanese rules is easily solved by refusing to play under
ridiculous rules. Yes, I do have strong opinions. :)

Álvaro.



On Tue, Mar 21, 2017 at 7:36 PM,  wrote:

>
> Why can't you reuse the same self played games but score them with a
> different komi value ? The policy network does not use the komi to choose
> its moves so it should make no difference.
>
>
> > On 21/03/2017 21:08, David Ongaro wrote:
> >>> But how would you fix it? Isn't that you'd need to retrain your value
> >>> network from the scratch?
> >>
> >> I would think so as well. But I some months ago I already made a
> >> proposal in this list to mitigate that problem: instead of training a
> >> different value network for each Komi, add a “Komi adjustment† value
> >> as
> >> input during the training phase. That should be much more effective,
> >> since the “win/lost† evaluation shouldn’t change for many (most?)
> >> positions for small adjustments but the resulting value network (when
> >> trained for different Komi adjustments) has a much greater range of
> >> applicability.
> >
> > The problem is not the training of the network itself (~2-4 weeks of
> > letting a program someone else wrote run in the background, easiest
> > thing ever in computer go), or whether you use a komi input or a
> > separate network, the problem is getting data for the different komi
> > values.
> >
> > Note that if getting data is not a problem, then a separate network
> > would perform better than your proposal.
> >
> > --
> > GCP
> > ___
> > 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
>
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] Zen lost to Mi Yu Ting

2017-03-21 Thread uurtamo . 
I guess that 1 point in such a game matters to the evaluation function.
Pretty fascinating. Can you not train for the two different rulesets and
just pick which at the beginning? Ignoring Chinese versus Japanese, just
training on komi? Or is the problem of Japanese rules the whole issue? (I.e
not komi)?

On Tuesday, March 21, 2017, Hideki Kato  wrote:

> The value network has been trained with Chinese rules and 7.5
> pts komi.  Using this for Japanese and 6.5, there will be some
> error in close games.  We knew this issue and thought such
> chances would be so small that postponed correcting (not so
> easy).
>
> Best,
> Hideki
>
> Pawe  Morawiecki: <
> caksbshpvd34hvjt-b+x73rdpg5-4wsxoezykbheslprewci...@mail.gmail.com
> >:
> >Hi,
> >
> >After an interesting game DeepZen lost to Mi Yu Ting.
> >Here you can replay the complete game:
> >http://duiyi.sina.com.cn/gibo_new/live/viewer.asp?sno=13
> >
> >According to pro experts, Zen fought really well, but it seems there is
> >still some issue how Zen (mis)evaluates its chances. At one point it
> showed
> >84% chance of winning (in the endgame), whereas it was already quite clear
> >Zen is little behind (2-3 points).
> >
> >Regards,
> >Pawel
> > inline file
> >___
> >Computer-go mailing list
> >Computer-go@computer-go.org 
> >http://computer-go.org/mailman/listinfo/computer-go
> --
> Hideki Kato >
> ___
> 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] Zen lost to Mi Yu Ting

2017-03-21 Thread Chun Sun 
How does Zen know it's playing a Japanese rule game? Can it be set to play
a Chinese rule game and hope to converge at the end?

On Mar 21, 2017 8:03 AM, "Hideki Kato"  wrote:

> The value network has been trained with Chinese rules and 7.5
> pts komi.  Using this for Japanese and 6.5, there will be some
> error in close games.  We knew this issue and thought such
> chances would be so small that postponed correcting (not so
> easy).
>
> Best,
> Hideki
>
> Pawe  Morawiecki:  4wsxoezykbheslprewci...@mail.gmail.com>:
> >Hi,
> >
> >After an interesting game DeepZen lost to Mi Yu Ting.
> >Here you can replay the complete game:
> >http://duiyi.sina.com.cn/gibo_new/live/viewer.asp?sno=13
> >
> >According to pro experts, Zen fought really well, but it seems there is
> >still some issue how Zen (mis)evaluates its chances. At one point it
> showed
> >84% chance of winning (in the endgame), whereas it was already quite clear
> >Zen is little behind (2-3 points).
> >
> >Regards,
> >Pawel
> > inline file
> >___
> >Computer-go mailing list
> >Computer-go@computer-go.org
> >http://computer-go.org/mailman/listinfo/computer-go
> --
> Hideki Kato 
> ___
> 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] Zen lost to Mi Yu Ting

2017-03-21 Thread cazenave 

Why can't you reuse the same self played games but score them with a
different komi value ? The policy network does not use the komi to choose
its moves so it should make no difference.


> On 21/03/2017 21:08, David Ongaro wrote:
>>> But how would you fix it? Isn't that you'd need to retrain your value
>>> network from the scratch?
>>
>> I would think so as well. But I some months ago I already made a
>> proposal in this list to mitigate that problem: instead of training a
>> different value network for each Komi, add a “Komi adjustment” value
>> as
>> input during the training phase. That should be much more effective,
>> since the “win/lost” evaluation shouldn’t change for many (most?)
>> positions for small adjustments but the resulting value network (when
>> trained for different Komi adjustments) has a much greater range of
>> applicability.
>
> The problem is not the training of the network itself (~2-4 weeks of
> letting a program someone else wrote run in the background, easiest
> thing ever in computer go), or whether you use a komi input or a
> separate network, the problem is getting data for the different komi
> values.
>
> Note that if getting data is not a problem, then a separate network
> would perform better than your proposal.
>
> --
> GCP
> ___
> 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] Zen lost to Mi Yu Ting

2017-03-21 Thread Gian-Carlo Pascutto 
On 21/03/2017 21:08, David Ongaro wrote:
>> But how would you fix it? Isn't that you'd need to retrain your value
>> network from the scratch?
> 
> I would think so as well. But I some months ago I already made a
> proposal in this list to mitigate that problem: instead of training a
> different value network for each Komi, add a “Komi adjustment” value as
> input during the training phase. That should be much more effective,
> since the “win/lost” evaluation shouldn’t change for many (most?)
> positions for small adjustments but the resulting value network (when
> trained for different Komi adjustments) has a much greater range of
> applicability.

The problem is not the training of the network itself (~2-4 weeks of
letting a program someone else wrote run in the background, easiest
thing ever in computer go), or whether you use a komi input or a
separate network, the problem is getting data for the different komi values.

Note that if getting data is not a problem, then a separate network
would perform better than your proposal.

-- 
GCP
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

[Computer-go] Different Rules

2017-03-21 Thread Ingo Althöfer 
Hi,

now we see how clever the DeepMind team was (and likely still is).
In both matches (against Fan Hui and Lee Sedol) Chinese rules
were applied.



Some years ago I performed experiments with Monte Carlo search 
in special non-zero sum games (with two players). The rules were made
in such a way that outcomes were possible that both sides were
winning according to their respective rules.
(An example from the Go framework: Black might think that komi is 
5.5 points, whereas White might think that komi is 7.5 points.)
RATHER OFTEN the outcome was a score where both sides thought
to have won. In the 5.5/7.5 komi example from Go  this means that
outcomes with +6 or +7 points for Black on the board would occur
often. 

Of course, this is not welcome for zero-sum games. But it is a hint
that in reallife scenarios (with non-zero-sum payoffs) Monte Carlo
heuristics (with their tendency to produce narrow wi0ns) might be
helpful in finding good compromises.

Ingo.
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] Zen lost to Mi Yu Ting

2017-03-21 Thread Paweł Morawiecki 
Hideki,

 Using this for Japanese and 6.5, there will be some
>> error in close games.  We knew this issue and thought such
>> chances would be so small that postponed correcting (not so
>> easy).
>>
>
But how would you fix it? Isn't that you'd need to retrain your value
network from the scratch?

Regards,
Paweł


>
> Oh, so that's why! Good luck with Zen's next two games.
>
> Aja
>
>
>> Best,
>> Hideki
>>
>> Pawe  Morawiecki: > wsxoezykbheslprewci...@mail.gmail.com>:
>> >Hi,
>> >
>> >After an interesting game DeepZen lost to Mi Yu Ting.
>> >Here you can replay the complete game:
>> >http://duiyi.sina.com.cn/gibo_new/live/viewer.asp?sno=13
>> >
>> >According to pro experts, Zen fought really well, but it seems there is
>> >still some issue how Zen (mis)evaluates its chances. At one point it
>> showed
>> >84% chance of winning (in the endgame), whereas it was already quite
>> clear
>> >Zen is little behind (2-3 points).
>> >
>> >Regards,
>> >Pawel
>> > inline file
>> >___
>> >Computer-go mailing list
>> >Computer-go@computer-go.org
>> >http://computer-go.org/mailman/listinfo/computer-go
>> --
>> Hideki Kato 
>> ___
>> 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
>
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] Zen lost to Mi Yu Ting

2017-03-21 Thread Aja Huang via Computer-go 
On Tue, Mar 21, 2017 at 10:48 AM, Hideki Kato 
wrote:

> The value network has been trained with Chinese rules and 7.5
> pts komi.  Using this for Japanese and 6.5, there will be some
> error in close games.  We knew this issue and thought such
> chances would be so small that postponed correcting (not so
> easy).
>

Oh, so that's why! Good luck with Zen's next two games.

Aja


> Best,
> Hideki
>
> Pawe  Morawiecki:  4wsxoezykbheslprewci...@mail.gmail.com>:
> >Hi,
> >
> >After an interesting game DeepZen lost to Mi Yu Ting.
> >Here you can replay the complete game:
> >http://duiyi.sina.com.cn/gibo_new/live/viewer.asp?sno=13
> >
> >According to pro experts, Zen fought really well, but it seems there is
> >still some issue how Zen (mis)evaluates its chances. At one point it
> showed
> >84% chance of winning (in the endgame), whereas it was already quite clear
> >Zen is little behind (2-3 points).
> >
> >Regards,
> >Pawel
> > inline file
> >___
> >Computer-go mailing list
> >Computer-go@computer-go.org
> >http://computer-go.org/mailman/listinfo/computer-go
> --
> Hideki Kato 
> ___
> 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] Zen lost to Mi Yu Ting

2017-03-21 Thread Hideki Kato 
The value network has been trained with Chinese rules and 7.5 
pts komi.  Using this for Japanese and 6.5, there will be some 
error in close games.  We knew this issue and thought such 
chances would be so small that postponed correcting (not so 
easy).

Best,
Hideki

Pawe Morawiecki: 
:
>Hi,
>
>After an interesting game DeepZen lost to Mi Yu Ting.
>Here you can replay the complete game:
>http://duiyi.sina.com.cn/gibo_new/live/viewer.asp?sno=13
>
>According to pro experts, Zen fought really well, but it seems there is
>still some issue how Zen (mis)evaluates its chances. At one point it showed
>84% chance of winning (in the endgame), whereas it was already quite clear
>Zen is little behind (2-3 points).
>
>Regards,
>Pawel
> inline file
>___
>Computer-go mailing list
>Computer-go@computer-go.org
>http://computer-go.org/mailman/listinfo/computer-go
-- 
Hideki Kato 
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] What a week ...

2017-03-21 Thread Nick Wedd 
On 21 March 2017 at 03:45, "Ingo Althöfer" <3-hirn-ver...@gmx.de> wrote:

> * Will AlphaGo accept the challenge by FineArt?
> (or Google the one by 10Cent, respectively ?)
>
> * Will DeepZenGo win the World Championships, starting right now?
>
> * Will some Robin Good rise from the underwoods to
> save humanity's honour?
>
> * Which of the currently three top bots will show up in the
> European Go Congress in Oberhof in July/August 2017?
>

Will there be a bot event in Oberhof?  I can see no mention of one at
http://www.egc2017.eu/events/ . Admittedly the schedule is very hard to
read, and resists attempts to enlarge it with ctrl-+.

Cheers,

Nick


>
> Questions over questions ...
>
> **
> And one unasked answer:
> TenCent was founded in November 1998 with the help of
> US-american risc capital.
>


> Cheers, Ingo (pressing his thumbs for DeepZenGo right now).
> ___
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go




-- 
Nick Wedd  mapr...@gmail.com
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] What a week ...

2017-03-21 Thread Detlef Schmicker 

> * Which of the currently three top bots will show up in the 
> European Go Congress in Oberhof in July/August 2017?

just set up one of the top open source bots:

on moderate hardware

ray: http://www.dragongoserver.net/userinfo.php?uid=97868

and if this is too strong for europe

oakfoam: http://www.dragongoserver.net/userinfo.php?uid=97704
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go