G
o
-
r
e
l
a
t
e
d
R
e
s
e
a
r
c
h
a
t
t
h
e
U
n
i
v
e
r
s
i
t
y
of A
l
b
e
r
t
a
Martin M lerDepar加entof Computing Science University of Alberta
Edmon旬n, C佃ada mmueller@cs.国lbe此a.ca
A
b
s
t
r
a
c
t
We give a brief outline and referencesto 血erecentGoュ rela旬dresearch which was performed 泊 ourgroup at the University of Alberta inEdmonωn, C姐ada.
Key words: Go
,
game 回e search, cOß也inatorial game theoryInr,回entyears
,
our work hasfocused on exactso・lutionsωsubproblems of Go
,
asopposedωheuristicapproachesωthe whole game. A 明or c町'ent in節子 est is how to apply these methods in realg創neplay. In
也efollowing sections we will briefly describe the pas
.
t
present andfu細胞 pl阻sin 血eseareas.1
Games R
e
s
e
a
r
c
h
a
t
t
h
e
U
n
i
v
e
r
-
3
r
Tsume・Goi
e
s
and S
a
f
e
t
y
o
f
Territo・s
i
t
y
o
f
A
l
b
e
r
t
a
The University of Albertaishome 旬 oneof the largest games T事searchgroups in the world. We do research inm組y games 阻dgame-related fields
,
suchωPoker, checkers, Othello, commercialg佃les,阻d AI planュ ning.Inform甜on about the projects currently under way can be found at http://www . cs. ualberta. cargames/. Thisextended め紺'llCt briefly outュ lines thoseprojects 也at are related to the game of Go and provides a list of references for further readュ ing on theωpics mentionedin 也e 匂lk. The home pageof 血ecomputer Go groupishttp://www . cs . ualberta.ca/ 陶games/go/.2 G
o
-
P
l
a
y
i
n
g
Programs
Member冨 ofour group are developing two Go pro・ grams: Markus Enzenberger'sN
e
u
r
o
G
o
[1] is based on a neural network in combinationwi血 search. It achieved excellent results on the 9 x 9 board Exp
l
o
r
e
r
[6.7]isa more 回ditionalprogram based on knowledge and search.Itw鎚 originallydeveloped by theau白or 阻disnow being used by several group members for theirr白earch.Work on Tsume-Go and the special case ofthe one-eye problem [4] will be described in detail in the talk by Kishimoto.
P
r
ovingthe 鈍fetyof territoriesisac
1
0sely relatedωpic. The solver developed during Xiaozhen Niu's MSc thesis research [11] greatly improves upon my previous solver,
which was described at GPW in1997.
4 C
o
m
b
i
n
a
t
o
r
i
a
l
Games and Go
Endgames
Combinatorialg:創ne theory is a divide-and-
c
onquer search model which issuitable for Go endgames.T
e
m
ュ
p
e
r
a
t
u
r
e
D
i
s
c
o
v
e
r
y
S
e
a
r
c
h
[9Jisthe first forward search algori由m ableωcompute or approximate the mean andtemperaωre of 組創bi回ryloopfreecombinaω,rial game.Inexperiments on sums of Amazons endgames,
itco町ic時ly outpe由rms 向Il・,boardαβsearch. In [10],
we s飽rted research 皿ωh仰旬蹴 searchω achieveefficient 姐d precisealgori也ms for playing sums ofhotg創nes.C
o
n
d
i
t
i
o
n
a
l
C
o
m
b
i
n
a
t
o
r
i
a
l
Games
[8] are a framework 伽textendscombinaωrial g:佃le theory by modeling dependencies between subgames.n L n L
They are areexpres鈎d 鎚 nonlocalconditions that deュ ten凶newhether moves are legal.
5 S
e
a
r
c
h
T
e
c
h
n
i
q
u
e
s
In his PhD research, Kishimoωbas ex'飽nded 必:'pnto deal with cycles as0∞町並 00 [4]. He developed a general and efficientsolutionω 白egraph history tnュ terac,陶'nproblem(OHI) [2
,
3,
5]. This work is used泊 00as weIIas 泊姐 ongoing a恥mpt ぬ solve 也e 伊me of checkers http://www . cs. ualberta. ca;-chinook/.
6 C
u
r
r
e
n
t
Work and F
u
t
u
r
e
P
l
a
n
s
Some oftheω,pics that we are currentlyinter関tedin are:
Higb-IevelPlan凶,ng 泊Go We areinvesti酔伽.gmod els such as HTN (Hierarcbical11間kNetwork) in
0
0
.
Oame-SAT [12Jis 姐 abs回ctmodel of deュ pendencies between games.Reinfortement Learning in Go In conjunction with Prof.
R
i
cbSutめnat 也e Unive悶ityofA1be由, weare studying new appro釦bes for 田泊g le鉱山,gin
0
0
.
Dependenty Analysis We w叩tωdevelop a bet蹴 model for dealing with dependent subgoals in
0
0
.
One example are endgames wbich are not comュ pletelyse戸rated.R
e
f
e
r
e
n
c
e
s
[1 J M. Enzenberger. Evaluation in
00
by a neural netュ work usingso食 segmen刷on.In J. van den Herik, H. Iida,姐dE. Heinz, ediωrs, Advanα沼 inComュ puter Games 10,
pages 97 -108.Kl
uwer,
2004. [2] A.Kishlmoto 佃dM. Mler. A solu姐onω 血eOHI problem fordepth-命stpr,∞f-numbersearcb. In Proceedings
0
1
the 7thJotnt 白ゆm脚 onln lormationScienc町 Jα's2003,
p昭es489 -492,
2ω3.
[3] A.Kishlmoto 佃dM. M鶴Ier. A solutionω 血e OHI problem fordepth-面stpr,∞f-numbersearcb
,
2∞3. 19 pages.AcCel蜘d 12/2∞3 forInfonna・姐onSciences.
[4] A.Kishlmoto 姐dM. Mler. Df-pn 泊 00: 姐
applicationωthe one喝:yeproblem. In J.
v
a
n
den Herik, H.Iida,個dE. Heinζediωrs, Advanc,倒的 Computer Gam四 10,pages 125 -141
.
KIuw町2∞4.
[5] A.Kisbimoωand M. Mler. A general soluュ tionω 由egrapbbisωry interaction problem. In Nili防御~nthNationalCo,ゆm舵~eon A
r
t
V
ctal Inュtellige脱却e (AAAJ 20例), pages 併4-649,SanJose, CA, 2ω4.
[6] M. Mler. Co,岬uter Go
a
s
a Sum0
1
Local Games: An Application01 ω'mbina,ωrlal Game mωry. PbD thesis,
ETH Zürich, 1995. Diss. ETH Nr.l1
.
0
0
6
.
[7] M. Mler. Counting the score: Positionevalua・
説on 泊 compu飽rOo.ICGA ./I.側rnal,25(4):21
9-228, 2ω2. [8] M. Mler. Conditional combinatorialgames,個d tbeir app
I
i
cation to analyzingc叩旬血graces in0
0
.
InformationScieru耽 154(3-4):189-202,
2∞3. 伊] M. MüIIer,
M. Enzenberger,姐d J. Scbaeffer. Temperature discovery searcb.In Nineteenth Naュ陶'nal 白砂'renæon ArtificiallnteU取næ(AAAI
20,何人 pages65ふ663, S削 Jωe, CA, 2ω4. [10] M. M鶴Ier and Z. Li. Locally informed global
総arcb for sums ofcombina旬,rial 伊mes, 2ω4. 11 pages. Accepted4/2似)4 forCompu旬お佃d Games 2似)4, Ramat
-Gan,
I
s
r
a
e
l.To 句'pt'af泊 Springer Ver
I
agLec伽re Notes 泊 COII司?uterS
c
iュ ence (LNCS).[
1
1
]
X
.
Niu 組dM. Mler. An improved 純fety 回lver forcomp脚r00
,
2ω4. 16P略es. A,∞:ep旬d 4/2ω4 for Computers and Games 2'似>4, Ramat.・O
an,
Israel.To 叩伊釘泊 Sp由gerVerlag L民tureNo旬S 凶 Computer
S
c
ience(LNCS).[12] L.Zbao 組dM.M剖ler. 0卸le-SAT: A prel泊ト inaryrepo
r
t
.
In SeventhInter聞tional 白砂F en on Theory and Appli制御nsolS州陶btlity 及sting(SAT 20,叫ん P唱es357-362,
Vancouver,
Canad