Page 1 of 1

Capture go in Hactar

Posted: Tue Jan 01, 2013 11:57 am
by yeellow
Thanks. It is a bit strange that your server is weaker in capture go than in normal go, I though that capture go should be easier for computers...

After last update (or the one before) I can no longer run intermediate or expert levels nor capture go 3 or 5. I think it is a bug (I can play capture go 3 and 5 on my tablet, where I previously played some games of that type).

Best regards,
Marcin

Re: Hactar 1.3 versions

Posted: Tue Jan 01, 2013 8:29 pm
by lpaatero
Hello,

The capture go engine is completely different from go engine, they have nothing common.
Both are monte-carlo engines, but heuristics are quite different.

In particular capture-go simulations suffer from problem of playing just one stone into opponents territory, and then losing when it is captured.
In go this kind of single stone never really matters, but in capture go it is critical. So in essence this add irrelevant "noise" to any long simulations.
It looks possible to create heuristics to reduce these moves, but it will take time.

For non-working levels, can you look up number in about hactar (in main screen options) and let me know it?
Key number is the number in parantheses after "version: 1.3.5"

Edit: in last two days I have fixed some problems in server side. These could have created *temporary* problem in playing go, but should not impact capture go. And now there should not be any problems.

regards
Lauri

Re: Hactar 1.3 versions

Posted: Thu Jan 03, 2013 5:11 pm
by yeellow
On my tablet I have version 1.3.6 and capture 3 and capture 5 are grayed out. On my phone they are active, but I remember I played there on both those levels, maybe it matters. All other levels are active on my tablet (Asus prime).

I agree that some heuristics would be different - keeping solid group is much more important and sacrifice is not an option. Also a group with 2 space eye lives... It's interesting that small change in rules makes such different game.

You could also think about making board 10x10 for capture go (if it is easy), or maybe 8x8 would make computer stronger? Even number of dimensions makes it more symmetric.

Another issue: how can I play handicap go or change komi when playing normal go?

I've also noticed that hactar wrongly decided that my group was alive in 9x9 normal go game. I noticed that during counting it is not possible to dispute group status and I was surprised to win the game with a dead group. You can find sgf pasted below (check black group at the top - for Hactar it is alive and black wins):

(;AP[GOWrite:3.0.2]FF[4]GM[1]CA[UTF-8]SZ[9]ST[2]DT[2013-01-02]GWRICI[104a5ab856642a9]PW[Hactar]KM[6.5]GN[ ]PB[yeellow]RE[b+26.5]GWRIGI[a54c2957-39d9-45fe-bf9e-ff0b48df383d]PM[2]GWRIEC[w]FG[259:]RU[Japanese]GWRIEL[GO1-09@0]
;B[cf];W[fe];B[eg];W[fg];B[gc];W[ef];B[df];W[dg];B[eh];W[dh];B[fh];W[gh];B[gg];W[gf];B[hg];W[hh];B[ff];W[ee];B[hf];DD[ff]W[fg];B[ch];W[he];B[ff];W[ge];B[fg];W[dd];B[eb];W[db];B[ec];W[dc];B[hd];W[id];B[ic];W[ie];B[bc];W[hb];B[hc];W[gb];B[fc];W[ea];B[fa];W[da];B[fb];W[ib];B[ha];W[bd];B[be];W[bb];B[cd];W[cc];B[ad];W[di];B[ci];W[ab];B[ac];DD[aa]W[ba];B[de];DD[aa]W[if];B[ig];DD[aa]W[fi];B[cg];DD[aa]W[ed];B[ia];DD[aa]W[tt];B[tt]
;TB[ae][af:bi][bd][ce][dg:dh][di:ii][ga][gb:ib][gh:ih]TW[aa][ca:cb]
)

Re: Capture go in Hactar

Posted: Thu Jan 31, 2013 7:44 pm
by lpaatero
Sorry for delay, for some reason I did not get notification from forum about your posting.

Currently capture go results are not synchronized to server.
As games are fully local, I does not happen "automatically".

----

I wonder if anyone claiming capture go to be easier has actually programmed one ever;)

I did do bit of work on capture go strenght, and it seems that AMAF (all-moves-as-first) heuristis is not as useful as I believed.
Most of simulations end long before filling the board. Currently engine is really bad only at the beginning. If it gets past it, it can play quite good endgame. Have to think about opening library at some point later.

One big practical problem there is: I do not know any other atari-go program. So I can tune parameters only against itself and human. Self tuning does not lead to very good results (learns to trick itself).

----

The game ending is really weird. In order to pass in this position, Hactar must believe it cannot play remaining dames. And as result, black group is alive. I cannot see any reason for this. Sometimes Hactar gets sekis wrong, this is known (but complex) problem.

EDIT: I found the log for the game. It seems engine calculated end position correcly (3 points for white). Unfortunately I do not have log of sever events anymore, so I cannot determine where / why this has gone wrong.

I have not included agreement phase to game ending. In most cases it would be just distraction, and beginner would not know what to do with it (except to "cheat" a win).
I have not included arbitrary handicaps to keep things simple and straight forward. In most cases people want to even games, so I tried to maximize their amount.