Fuego

> Hi devel,
> 
> i played several 9x9 games against FuegoBot2 at kgs go server.
> 
> it is obvious to me (despite the manual) that the late endgame of Fuego is very bad.
> i won several games cause of this buggy behaviour. (account laska0, you may download the games (where Fuego resigned or lost by a small margin))
> 
> does anyone work on this problem? (please don't tell me it does not exist, like the manual.)
> being a developer myself, i'd like to help to fix this problem.
> 
> best
>         vim

Hi vim,

I looked at some of your recent games. I did not see any where bad endgame caused a loss. Can you give us the specific games please?

The program does play silly moves after it is already lost in many games. This is typical of Monte Carlo engines, which only try to win and do not care by how much they lose. Still, I agree it should be fixed so it keeps playing normal-looking moves. Your help in this regard is very welcome. Maybe some form of dynamic komi would work for that.

I also found a serious problem with Fuego not understanding nakade in one of the games. I have added it as
regression/sgf/games/2011/KGS/FuegoBot2-laska0-3.sgf
and made a test case. It is an interesting seki position where Fuego can even play tenuki, but the move it played kills its own group. I ran it a few times, and it sometimes prefers D2, but more often plays G1 as in the game, which is suicidal. I suspect that something goes systematically wrong in the playouts.

	Martin

Nice find.  You can actually hope for an ELO gain when you fix such nasty
plays.


On Thu, Apr 21, 2011 at 1:48 PM, Martin Mueller <mmueller@...:

>
> > Hi devel,
> >
> > i played several 9x9 games against FuegoBot2 at kgs go server.
> >
> > it is obvious to me (despite the manual) that the late endgame of Fuego
> is very bad.
> > i won several games cause of this buggy behaviour. (account laska0, you
> may download the games (where Fuego resigned or lost by a small margin))
> >
> > does anyone work on this problem? (please don't tell me it does not
> exist, like the manual.)
> > being a developer myself, i'd like to help to fix this problem.
> >
> > best
> >         vim
>
> Hi vim,
>
> I looked at some of your recent games. I did not see any where bad endgame
> caused a loss. Can you give us the specific games please?
>
> The program does play silly moves after it is already lost in many games.
> This is typical of Monte Carlo engines, which only try to win and do not
> care by how much they lose. Still, I agree it should be fixed so it keeps
> playing normal-looking moves. Your help in this regard is very welcome.
> Maybe some form of dynamic komi would work for that.
>
> I also found a serious problem with Fuego not understanding nakade in one
> of the games. I have added it as
> regression/sgf/games/2011/KGS/FuegoBot2-laska0-3.sgf
> and made a test case. It is an interesting seki position where Fuego can
> even play tenuki, but the move it played kills its own group. I ran it a few
> times, and it sometimes prefers D2, but more often plays G1 as in the game,
> which is suicidal. I suspect that something goes systematically wrong in the
> playouts.
>
>        Martin
>
> ------------------------------------------------------------------------------
> Fulfilling the Lean Software Promise
> Lean software platforms are now widely adopted and the benefits have been
> demonstrated beyond question. Learn why your peers are replacing JEE
> containers with lightweight application servers - and what you can gain
> from the move. http://p.sf.net/sfu/vmware-sfemails
> _______________________________________________
> fuego-devel mailing list
> fuego-devel@...
> https://lists.sourceforge.net/lists/listinfo/fuego-devel
>

Hi Martin,

thx for the trouble investigating the games ...

On Thu, Apr 21, 2011 at 7:48 PM, Martin Mueller <mmueller@...:

>
> Can you give us the specific games please?
>
hhhmmm, i should, i guess ... wait ...


>
> The program does play silly moves after it is already lost in many games.
> This is typical of Monte Carlo engines, which only try to win and do not
> care by how much they lose.

sorry, but, NO, it is not that.
it throws away a sure win.
some of the games even had observers, seeing me lose and they all asked
"why" after Fuego played this ********.



> Still, I agree it should be fixed so it keeps playing normal-looking moves.
> Your help in this regard is very welcome. Maybe some form of dynamic komi
> would work for that.
>
i do not care about "random" moves, after losing, or building concrete walls
and winning with .5 points.
again: it's before losing


I also found a serious problem with Fuego not understanding nakade in one of
> the games. I have added it as
> regression/sgf/games/2011/KGS/FuegoBot2-laska0-3.sgf
> and made a test case. It is an interesting seki position where Fuego can
> even play tenuki, but the move it played kills its own group. I ran it a few
> times, and it sometimes prefers D2, but more often plays G1 as in the game,
> which is suicidal.




> I suspect that something goes systematically wrong in the playouts.
>
yes, that's the point
(as Micheal does, i even guess, if the problem can be fixed, the rating will
increase quite a bit.)

It's Chinese rules.  Fuego wins.



On Fri, Apr 22, 2011 at 2:33 PM, Lambert Herrgesell
<zero815@...:

> Hi,
>
> ok, i did a download of the latest version (Fuego 1.1) and played some
> games.
> well, at a first glance, it feels much better as the version 0.4.1 playing
> on kgs
> (but, maybe i'm mistaken by just a label ... ;) )
>
> i guess, it has something to do with kos.
>
> attached is a file, where at move 82 (or so) Fuego leads by 1.5 points
> if i do not count wrong:
> b (me) has 23 points at that time, 12 terri + 11 captured
> w (fuego) 18 = 9 terri + 9 captured
> with 6.5 komi fuego should win by 18+6.5 = 24.5 to 23 points.
> but then it goes berserk and i get a present of 7 (!) points until it
> resigns.
>
> (please excuse my exceptionally bad go.)
>
> best
>           vim
>
>
>
>
> ------------------------------------------------------------------------------
> Fulfilling the Lean Software Promise
> Lean software platforms are now widely adopted and the benefits have been
> demonstrated beyond question. Learn why your peers are replacing JEE
> containers with lightweight application servers - and what you can gain
> from the move. http://p.sf.net/sfu/vmware-sfemails
> _______________________________________________
> fuego-devel mailing list
> fuego-devel@...
> https://lists.sourceforge.net/lists/listinfo/fuego-devel
>
>

Oops, that's not th case.  Sorry.

On Fri, Apr 22, 2011 at 2:44 PM, Michael Williams <
michaelwilliams75@...> wrote:

> It's Chinese rules.  Fuego wins.
>
>
>
>   On Fri, Apr 22, 2011 at 2:33 PM, Lambert Herrgesell <
> zero815@...> wrote:
>
>>  Hi,
>>
>> ok, i did a download of the latest version (Fuego 1.1) and played some
>> games.
>> well, at a first glance, it feels much better as the version 0.4.1 playing
>> on kgs
>> (but, maybe i'm mistaken by just a label ... ;) )
>>
>> i guess, it has something to do with kos.
>>
>> attached is a file, where at move 82 (or so) Fuego leads by 1.5 points
>> if i do not count wrong:
>> b (me) has 23 points at that time, 12 terri + 11 captured
>> w (fuego) 18 = 9 terri + 9 captured
>> with 6.5 komi fuego should win by 18+6.5 = 24.5 to 23 points.
>> but then it goes berserk and i get a present of 7 (!) points until it
>> resigns.
>>
>> (please excuse my exceptionally bad go.)
>>
>> best
>>           vim
>>
>>
>>
>>
>> ------------------------------------------------------------------------------
>> Fulfilling the Lean Software Promise
>> Lean software platforms are now widely adopted and the benefits have been
>> demonstrated beyond question. Learn why your peers are replacing JEE
>> containers with lightweight application servers - and what you can gain
>> from the move. http://p.sf.net/sfu/vmware-sfemails
>> _______________________________________________
>> fuego-devel mailing list
>> fuego-devel@...
>> https://lists.sourceforge.net/lists/listinfo/fuego-devel
>>
>>
>

Yes, it seems to be a clash between you playing Japanese rules and Fuego playing Chinese rules.
Fuego does support Japanese rules as well, but this support is incomplete.

Try setting
go_rules japanese

and see if it helps. If not, send us a bug report.

If you use GoGui, you can choose "score" in the game menu, and you will see both types of scoring.

	Martin

> It's Chinese rules.  Fuego wins.
> 
> 
>  
> On Fri, Apr 22, 2011 at 2:33 PM, Lambert Herrgesell <zero815@...> wrote:
> Hi,
> 
> ok, i did a download of the latest version (Fuego 1.1) and played some games.
> well, at a first glance, it feels much better as the version 0.4.1 playing on kgs
> (but, maybe i'm mistaken by just a label ... ;) )
> 
> i guess, it has something to do with kos. 
> 
> attached is a file, where at move 82 (or so) Fuego leads by 1.5 points
> if i do not count wrong:
> b (me) has 23 points at that time, 12 terri + 11 captured
> w (fuego) 18 = 9 terri + 9 captured
> with 6.5 komi fuego should win by 18+6.5 = 24.5 to 23 points.
> but then it goes berserk and i get a present of 7 (!) points until it resigns.
> 
> (please excuse my exceptionally bad go.)
> 
> best
>           vim
> 
> 
> 
> ------------------------------------------------------------------------------
> Fulfilling the Lean Software Promise
> Lean software platforms are now widely adopted and the benefits have been
> demonstrated beyond question. Learn why your peers are replacing JEE
> containers with lightweight application servers - and what you can gain
> from the move. http://p.sf.net/sfu/vmware-sfemails
> _______________________________________________
> fuego-devel mailing list
> fuego-devel@...
> https://lists.sourceforge.net/lists/listinfo/fuego-devel
> 
> 
> ------------------------------------------------------------------------------
> Fulfilling the Lean Software Promise
> Lean software platforms are now widely adopted and the benefits have been 
> demonstrated beyond question. Learn why your peers are replacing JEE 
> containers with lightweight application servers - and what you can gain 
> from the move. http://p.sf.net/sfu/vmware-sfemails_______________________________________________
> fuego-devel mailing list
> fuego-devel@...
> https://lists.sourceforge.net/lists/listinfo/fuego-devel

Hi Martin,

On Sat, Apr 23, 2011 at 11:47 PM, Martin Mueller <mmueller@...:

> Yes, it seems to be a clash between you playing Japanese rules and Fuego
> playing Chinese rules.
>
i don't care about the differences here, and please don't push it aside so
easy.

--> it kills it's groups, it gives me prisoners. <---

(and not something about dame, i am to stupid to understand.)

if you don't or can't see, ok.

best
        vim

On 2011-04-24, at 12:33 AM, Lambert Herrgesell wrote:

> Hi Martin,
> 
> On Sat, Apr 23, 2011 at 11:47 PM, Martin Mueller <mmueller@...> wrote:
> Yes, it seems to be a clash between you playing Japanese rules and Fuego playing Chinese rules.
> i don't care about the differences here, and please don't push it aside so easy.
> 
> --> it kills it's groups, it gives me prisoners. <---
> 
> (and not something about dame, i am to stupid to understand.)
> 
> if you don't or can't see, ok.

Prisoners do not count under Chinese rules. Try score in GoGui and you can see both ways of counting.

    Martin

Hi Rich,
yes I am aware there is a bug with seki in the safety solver.
fuego/regression/sgf/seki/static-safety-seki-bug.sgf is a similar case.
See static_safety_seki.tst, cases 40-43.

I promise to fix it this summer :)

I have added your test case in the same space, and also added tests for the first time the seki is created in this game.


	Martin

> I started looking at this game and ran into what I think is a bug in the safety solver. Unfortunately, I do not think fixing this bug will help Fuego play better in this situation. The bug could affect Fuego's endgame play. The attached SGF is one possible extension of FuegoBot2-lasko0-3.sgf where the seki in the bottom is realized. At move 67, the static safety solver incorrectly labels the bottom as safe for black. Martin, could you look into this?