[Computer-go] KGS tournament rules

Aja Huang ajahuang at google.com
Fri Jun 5 04:15:44 PDT 2015


On Thu, Jun 4, 2015 at 9:56 PM, Nick Wedd <maproom at gmail.com> wrote:

> I have been asked
>
> Your page [ http://www.weddslist.com/kgs/rules.html ] says:
>>
>> All the code in it that is in any way involved in move-generation (i.e.
>>> anything that causes the program to prefer one move to another) or position
>>> evaluation must be unique among the entrants. Code that is involved only in
>>> non-essential parts of the program, such as input/output, or scoring the
>>> position after the game is over, need not be unique. If two or more people
>>> want to submit programs containing the same code, then the author of that
>>> code shall decide which may enter.
>>
>>
>> Would it be acceptable for me to use a (non-Go-specific) neural network
>> package that I didn't write?
>>
>
> My immediate inclination is to say "Yes. It's like using a compiler that
> you didn't write."  But
> I fear it may be more complicated than that.
>

I would suggest "Yes", too. Neural networks are just like computer
programs. They accept inputs and produce outputs. You can either write a
neural network yourself, buy one, or download one from somewhere. If you
train a neural network yourself, it doesn't matter which libraries you are
using, just as it doesn't matter in which programming language you write a
MCTS Go program.

I think what really matters is that there shouldn't be two identical neural
networks in a tournament. This raises another question: should we allow
non-open-source, or commercial neural networks? This is equivalent to ask:
should we allow people to buy Crazy Stone or Zen softwares and enter KGS
tournaments? I would suggest "No". Maybe we should have a policy something
like this:

1. For open source programs (including NN), allow at most one entry for
each program.
2. For non-open source programs (including NN), only the author can
register for the tournament.

Aja


> For now, the rule is that if you enter a KGS bot tournament using a neural
> net that you did not write, your entry will be accepted, buy you must
> specify what neural net you are using.
>
> But I would like to discuss the issue, and accept the consensus of this
> list.  I have never used a neural net, and my understanding of how they
> work is close to zero.  I naively imagine it goes like this:
>   1.  You obtain a neural net, by buying one, downloading a free one, or
> getting one from a colleague.
>   2.  You install it on your computer.
>   3.  You configure it by setting some parameters.
>   4.  You specify how its board state representation will work (I have
> very little idea about this).
>   5.  You train it, maybe by feeding it a large database of professional
> games.
>   6.  You test the results. Quite likely you realise it hasn't gone well,
> and redo from step 3.
>   7. You add a harness that attaches it to kgsGtp, and maybe to some other
> programs.
>
> I look forward to becoming better informed.  I know that if someone writes
> a praiseworthy program in say C, the creator of his C compiler will deserve
> and expect none of the credit. I suspect things may be different with
> neural nets.
>
> Nick
> --
> Nick Wedd      maproom at gmail.com
>
> _______________________________________________
> Computer-go mailing list
> Computer-go at computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://computer-go.org/pipermail/computer-go/attachments/20150605/f66fcad4/attachment.html>


More information about the Computer-go mailing list