[Computer-go] ManyFaces vs Aya today (round 8 of the slow bottournament)
dailey.don at gmail.com
Wed May 25 05:53:31 PDT 2011
On Wed, May 25, 2011 at 8:23 AM, <valkyria at phmp.se> wrote:
> I let Valkyria play from this position on an old single core P4. It plays
> out and wins the semeai even before it has to do so.
> So by demonstrating this I have shown that with modest computational effort
> Valkyria thinks it is pretty simple. Or maybe more correctly: Valkyria can
> see moves that makes the game very simple.
> So if Aya resigns a won position it means Aya needs to be improved not all
> go programs in general.
> Also just because Valkyria plays reasonably well in this position does not
> mean that Valkyria is strong. It just happen to have enough knowledge to
> handle the tactics with a narrow search in this very position. I am sure
> there are a lot of positions Aya will outplay Valkyria.
> Most MCTS programs used to be light in the sense close to buggy.
> This is not more the case and the programs are getting stronger and
> stronger. But the process of adding knowledge is a trial and error process
> and different programmers have done different stuff. Also there is the
> tradeoff between speed and specialized knowledge. I always programmed
> Valkyria with no fear of slowing it down ...
As a computer chess programmer I struggle with this concept a lot. Do I
fix the problem and weaken the program or do I accept the weakness. My
general approach if it's a close call is to fix any problems that are not
solved by a modest amount of additional depth. I believe that as program
look deeper and get stronger those kind of errors are going to stand out
more and more and must be fixed even if there is a small cost.
Computer go is still several years behind computer chess in many ways, but
quickly catching up. So my definition of what is "close call" may be
different than yours. In computer chess at the highest levels you might
spend a week to get a 2% faster programs or a 2 ELO gain (which could be 100
ELO in a year.) In computer Go it's more productive looking for big 20-30
ELO improvements. I still remember the days where some simple idea (the
proverbial 1 line change) could give you 20 or 30 ELO.
What you just said about no fear .. I remember John Stanback, the Zarkov
programmer, used to say that his evaluation was so slow that he could add
just about anything to it with no fear of slowing the program down
> and I make any change that I think is necessary to play correctly even if
> the tactics are rare cases. Hence it tends to do well in positions where
> some program has serious problems. Overall though this does not make
> Valkyria the strongest program.
> So my point is: for most non trivial tactical situations, some go
> programmers has a solution already. It is not hard to improve playouts. But
> it is very hard if not impossible to improve playouts for every possible
> special tactical position.
> The question is : how do we tackle tactical positions in a general sense
> without slowing down playouts too much?
> Semeai is a good example. Valkyria is good on simple cases close to the
> winning capture but there are an unending amount of complex semeai where
> this knowledge only have an effect deep down in the playouts. Some algorithm
> of online learning of good tactics seems to be necessary.
> Quoting Stefan Kaitschick <Stefan.Kaitschick at Hamburg.de>:
> It's childishly simple for a human, large eye against small eye.
>> Programmers need to find a way to make this a simple position for bots
>> Solving it with great computational effort isn't good enough.
>> Aya resigned in winnning position...
>>> But in last position, to save B13 string,
>>> kill B10.
>>> to kill B10, play D6.
>>> to play D6, kill B8.
>>> Semeai is really difficult.
>>> Hiroshi Yamashita
>>> ----- Original Message ----- From: "David Fotland" <
>>> fotland at smart-games.com>
>>> To: <computer-go at dvandva.org>; <nick at maproom.co.uk>
>>> Sent: Wednesday, May 25, 2011 12:34 PM
>>> Subject: [Computer-go] ManyFaces vs Aya today (round 8 of the slow
>>> In this game, there was a big semeai on the left side. The result was
>>>> a won
>>>> position for Aya, but both Aya and ManyFaces thought that ManyFaces had
>>>> (or perhaps that it was a semeai), so eventually Aya resigned before it
>>>> played out.
>>>> A lucky win for ManyFaces, and a position for the report.
> Computer-go mailing list
> Computer-go at dvandva.org
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Computer-go