gnugo-devel
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [gnugo-devel] Edmondton impressions


From: bump
Subject: Re: [gnugo-devel] Edmondton impressions
Date: Tue, 13 Aug 2002 12:27:02 -0700

> > Among the talks, I found Georg Santze's talk about Dreihirn
> > interesting. He is from Althofer's group. The Dreihirn consists
> > of a human moderator selecting moves from two computers. The
> > resulting agent may be stronger than either the human moderator
> > or either program. This is convincingly proved in both Chess
> > and Go. An alternative scheme has the moderator selecting
> > from the top n moves of a computer program. I think such an
> > experiment with GNU Go could be useful in giving an upper
> > bound for the strength of metamachine which attempts to
> > select the best of GNU Go's top two moves. If such an experiment
> > were successful we could try to improve metamachine's selection
> > of the top two moves using full board search at 2 ply or deeper.
> 
> Maybe I will try to play moderator and see whether I can give GNU Go
> plain some handicap.

I tried such an experiment for one game a couple of weeks ago
and was not pleased by the result.

One problem was that if the top move is overvalued then that move that
overvalued move will be proposed over and over again reducing the dreihirn's
effective options to only one move. For much of the following game W
wanted to play S5 before it was big enough.

I can think of some strategies to avoid this.

Here is the game. The dreihirn is W. At the end of the game W has a
dead group at the top.

Dan

(;GM[1]FF[3]
RU[Japanese]SZ[19]HA[0]KM[5.5]
PW[White]
PB[Black]
GN[White (W) vs. Black (B)]
DT[2002-08-01]
;B[pd];W[cp];B[dd];W[pp];B[jd]
;W[cf];B[ch];W[cc];B[dc];W[cd]
;B[de];W[cb];B[df];W[bf];B[eh]
;W[md];B[pf];W[ep];B[qn];W[mf]
;B[qk];W[mh];B[gq];W[nq];B[rp]
;W[qq];B[lq];W[rq];B[oh];W[cm]
;B[gc];W[eq];B[ck];W[li];B[lo]
;W[jg];B[go];W[hg];B[jn];W[ie]
;B[id];W[ij];B[oc];W[db];B[mc]
;W[lc];B[ld];W[kc];B[kd];W[mb]
;B[nc];W[ib];B[nb];W[lb];B[ma]
;W[eb];B[ga];W[fr];B[bk];W[hr];
B[ge];W[hd];B[hc];W[fn];B[he];
W[hm];B[if];W[jl];B[hn];W[ll];
B[an];W[gm];B[jp];W[ro];B[np];
W[mq];B[mp];W[lr];B[im];W[il];
B[rn];W[kq];B[jq];W[ok];B[lp];
W[kr];B[hq];W[pi];B[ir];W[qp];
B[gr];W[qg];B[nf];W[rj];B[pl];
W[qf];B[re];W[oi];B[rk];W[qj];
B[nh];W[ni];B[ek];W[ne];B[of];
W[bo];B[on];W[qe];B[qd];W[rf];
B[rd];W[hb];B[gb];W[fl];B[gj];
W[ii];B[jb];W[gi];B[fi];W[hj];
B[gh]
)








reply via email to

[Prev in Thread] Current Thread [Next in Thread]