<s>
Computer	B-Application
Go	I-Application
is	O
the	O
field	O
of	O
artificial	B-Application
intelligence	I-Application
(	O
AI	B-Application
)	O
dedicated	O
to	O
creating	O
a	O
computer	B-Application
program	I-Application
that	O
plays	O
the	O
traditional	O
board	O
game	O
Go	O
.	O
</s>
<s>
The	O
best	O
efforts	O
of	O
the	O
1980s	O
and	O
1990s	O
produced	O
only	O
AIs	B-Application
that	O
could	O
be	O
defeated	O
by	O
beginners	O
,	O
and	O
AIs	B-Application
of	O
the	O
early	O
2000s	O
were	O
intermediate	O
level	O
at	O
best	O
.	O
</s>
<s>
Professionals	O
could	O
defeat	O
these	O
programs	O
even	O
given	O
handicaps	O
of	O
10+	O
stones	O
in	O
favor	O
of	O
the	O
AI	B-Application
.	O
</s>
<s>
Many	O
of	O
the	O
algorithms	O
such	O
as	O
alpha-beta	B-Algorithm
minimax	I-Algorithm
that	O
performed	O
well	O
as	O
AIs	B-Application
for	O
checkers	O
and	O
chess	B-Application
fell	O
apart	O
on	O
Go	O
's	O
19x19	O
board	O
,	O
as	O
there	O
were	O
too	O
many	O
branching	O
possibilities	O
to	O
consider	O
.	O
</s>
<s>
Some	O
AI	B-Application
researchers	O
speculated	O
that	O
the	O
problem	O
was	O
unsolvable	O
without	O
creation	O
of	O
human-like	O
AI	B-Application
.	O
</s>
<s>
The	O
application	O
of	O
Monte	B-Application
Carlo	I-Application
tree	I-Application
search	I-Application
to	O
Go	O
algorithms	O
provided	O
a	O
notable	O
improvement	O
in	O
the	O
late	O
2000s	O
decade	O
,	O
with	O
programs	O
finally	O
able	O
to	O
achieve	O
a	O
low-dan	O
level	O
:	O
that	O
of	O
an	O
advanced	O
amateur	O
.	O
</s>
<s>
The	O
key	O
insight	O
proved	O
to	O
be	O
an	O
application	O
of	O
machine	O
learning	O
and	O
deep	B-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
DeepMind	B-Application
,	O
a	O
Google	B-Application
acquisition	O
dedicated	O
to	O
AI	B-Application
research	I-Application
,	O
produced	O
AlphaGo	B-Application
in	O
2015	O
and	O
announced	O
it	O
to	O
the	O
world	O
in	O
2016	O
.	O
</s>
<s>
AlphaGo	B-Application
defeated	I-Application
Lee	I-Application
Sedol	I-Application
,	O
a	O
9	O
dan	O
professional	O
,	O
in	O
a	O
no-handicap	O
match	O
in	O
2016	O
,	O
then	O
defeated	B-Application
Ke	I-Application
Jie	I-Application
in	I-Application
2017	I-Application
,	O
who	O
at	O
the	O
time	O
continuously	O
held	O
the	O
world	O
No	O
.	O
</s>
<s>
Just	O
as	O
checkers	O
had	O
fallen	O
to	O
machines	O
in	O
1995	O
and	O
chess	B-Application
in	O
1997	O
,	O
computer	B-Application
programs	I-Application
finally	O
conquered	O
humanity	O
's	O
greatest	O
Go	O
champions	O
in	O
20162017	O
.	O
</s>
<s>
DeepMind	B-Application
did	O
not	O
release	O
AlphaGo	B-Application
for	O
public	O
use	O
,	O
but	O
various	O
programs	O
have	O
been	O
built	O
since	O
based	O
on	O
the	O
journal	O
articles	O
DeepMind	B-Application
released	O
describing	O
AlphaGo	B-Application
and	O
its	O
variants	O
.	O
</s>
<s>
It	O
has	O
long	O
been	O
considered	O
a	O
difficult	O
challenge	O
in	O
the	O
field	O
of	O
artificial	B-Application
intelligence	I-Application
(	O
AI	B-Application
)	O
and	O
is	O
considerably	O
more	O
difficult	O
to	O
solve	O
than	O
chess	B-Application
.	O
</s>
<s>
Many	O
in	O
the	O
field	O
considered	O
Go	O
to	O
require	O
more	O
elements	O
that	O
mimic	O
human	O
thought	O
than	O
chess	B-Application
.	O
</s>
<s>
Prior	O
to	O
2015	O
,	O
the	O
best	O
Go	B-Application
programs	I-Application
only	O
managed	O
to	O
reach	O
amateur	O
dan	O
level	O
.	O
</s>
<s>
Prior	O
to	O
AlphaGo	B-Application
,	O
some	O
researchers	O
had	O
claimed	O
that	O
computers	O
would	O
never	O
defeat	O
top	O
humans	O
at	O
Go	O
.	O
</s>
<s>
It	O
introduced	O
an	O
influence	O
function	O
to	O
estimate	O
territory	O
and	O
Zobrist	B-Application
hashing	I-Application
to	O
detect	O
ko	O
.	O
</s>
<s>
In	O
April	O
1981	O
,	O
Jonathan	O
K	O
Millen	O
published	O
an	O
article	O
in	O
Byte	O
discussing	O
Wally	O
,	O
a	O
Go	O
program	O
with	O
a	O
15x15	O
board	O
that	O
fit	O
within	O
the	O
KIM-1	B-Device
microcomputer	O
's	O
1K	O
RAM	O
.	O
</s>
<s>
Bruce	O
F	O
.	O
Webster	O
published	O
an	O
article	O
in	O
the	O
magazine	O
in	O
November	O
1984	O
discussing	O
a	O
Go	O
program	O
he	O
had	O
written	O
for	O
the	O
Apple	B-Device
Macintosh	I-Device
,	O
including	O
the	O
MacFORTH	B-Application
source	O
.	O
</s>
<s>
AIs	B-Application
who	O
played	O
on	O
the	O
Internet	O
Go	O
Server	O
(	O
IGS	O
)	O
on	O
19x19	O
size	O
boards	O
had	O
around	O
2015	O
kyu	O
strength	O
in	O
2003	O
,	O
after	O
substantial	O
improvements	O
in	O
hardware	O
.	O
</s>
<s>
In	O
1998	O
,	O
very	O
strong	O
players	O
were	O
able	O
to	O
beat	O
computer	B-Application
programs	I-Application
while	O
giving	O
handicaps	O
of	O
25	O
–	O
30	O
stones	O
,	O
an	O
enormous	O
handicap	O
that	O
few	O
human	O
players	O
would	O
ever	O
take	O
.	O
</s>
<s>
There	O
was	O
a	O
case	O
in	O
the	O
1994	O
World	O
Computer	B-Application
Go	I-Application
Championship	O
where	O
the	O
winning	O
program	O
,	O
Go	O
Intellect	O
,	O
lost	O
all	O
three	O
games	O
against	O
the	O
youth	O
players	O
while	O
receiving	O
a	O
15-stone	O
handicap	O
.	O
</s>
<s>
In	O
2006	O
(	O
with	O
an	O
article	O
published	O
in	O
2007	O
)	O
,	O
Rémi	O
Coulom	O
produced	O
a	O
new	O
algorithm	O
he	O
called	O
Monte	B-Application
Carlo	I-Application
tree	I-Application
search	I-Application
.	O
</s>
<s>
However	O
,	O
computers	O
"	O
score	O
"	O
a	O
terminal	O
leaf	O
of	O
the	O
tree	O
by	O
repeated	O
random	O
playouts	O
(	O
similar	O
to	O
Monte	B-Algorithm
Carlo	I-Algorithm
strategies	O
for	O
other	O
problems	O
)	O
.	O
</s>
<s>
The	O
intuitive	O
objection	O
-	O
that	O
random	O
playouts	O
do	O
not	O
correspond	O
to	O
the	O
actual	O
worth	O
of	O
a	O
position	O
-	O
turned	O
out	O
not	O
to	O
be	O
as	O
fatal	O
to	O
the	O
procedure	O
as	O
expected	O
;	O
the	O
"	O
tree	B-Algorithm
search	I-Algorithm
"	O
side	O
of	O
the	O
algorithm	O
corrected	O
well	O
enough	O
for	O
finding	O
reasonable	O
future	O
game	O
trees	O
to	O
explore	O
.	O
</s>
<s>
Programs	O
based	O
on	O
this	O
method	O
such	O
as	O
MoGo	O
and	O
Fuego	O
saw	O
better	O
performance	O
than	O
classic	O
AIs	B-Application
from	O
earlier	O
.	O
</s>
<s>
In	O
2009	O
,	O
the	O
first	O
such	O
programs	O
appeared	O
which	O
could	O
reach	O
and	O
hold	O
low	O
dan-level	O
ranks	O
on	O
the	O
KGS	B-Application
Go	I-Application
Server	I-Application
on	O
the	O
19x19	O
board	O
.	O
</s>
<s>
In	O
2011	O
,	O
Zen	B-Application
reached	O
5	O
dan	O
on	O
the	O
server	O
KGS	O
,	O
playing	O
games	O
of	O
15	O
seconds	O
per	O
move	O
.	O
</s>
<s>
The	O
account	O
which	O
reached	O
that	O
rank	O
uses	O
a	O
cluster	O
version	O
of	O
Zen	B-Application
running	O
on	O
a	O
26-core	O
machine	O
.	O
</s>
<s>
In	O
2012	O
,	O
Zen	B-Application
beat	O
Takemiya	O
Masaki	O
(	O
9p	O
)	O
by	O
11	O
points	O
at	O
five	O
stones	O
handicap	O
,	O
followed	O
by	O
a	O
20-point	O
win	O
at	O
four	O
stones	O
handicap	O
.	O
</s>
<s>
In	O
2013	O
,	O
Crazy	B-Application
Stone	I-Application
beat	O
Yoshio	O
Ishida	O
(	O
9p	O
)	O
in	O
a	O
19×19	O
game	O
at	O
four	O
stones	O
handicap	O
.	O
</s>
<s>
The	O
2014	O
Codecentric	O
Go	O
Challenge	O
,	O
a	O
best-of-five	O
match	O
in	O
an	O
even	O
19x19	O
game	O
,	O
was	O
played	O
between	O
Crazy	B-Application
Stone	I-Application
and	O
Franz-Jozef	O
Dickhut	O
(	O
6d	O
)	O
.	O
</s>
<s>
Franz-Jozef	O
Dickhut	O
won	O
,	O
though	O
Crazy	B-Application
Stone	I-Application
won	O
the	O
first	O
match	O
by	O
1.5	O
points	O
.	O
</s>
<s>
In	O
October	O
2015	O
,	O
Google	B-Application
DeepMind	I-Application
program	O
AlphaGo	B-Application
beat	O
Fan	O
Hui	O
,	O
the	O
European	O
Go	O
champion	O
,	O
five	O
times	O
out	O
of	O
five	O
in	O
tournament	O
conditions	O
.	O
</s>
<s>
In	O
March	O
2016	O
,	O
AlphaGo	B-Application
beat	O
Lee	O
Sedol	O
in	O
the	O
first	O
three	O
of	O
five	O
matches	O
.	O
</s>
<s>
AlphaGo	B-Application
won	O
the	O
final	O
match	O
two	O
days	O
later	O
.	O
</s>
<s>
In	O
May	O
2017	O
,	O
AlphaGo	B-Application
beat	O
Ke	O
Jie	O
,	O
who	O
at	O
the	O
time	O
was	O
ranked	O
top	O
in	O
the	O
world	O
,	O
in	O
a	O
three-game	B-Application
match	I-Application
during	O
the	O
Future	B-Application
of	I-Application
Go	I-Application
Summit	I-Application
.	O
</s>
<s>
In	O
October	O
2017	O
,	O
DeepMind	B-Application
revealed	O
a	O
new	O
version	O
of	O
AlphaGo	B-Application
,	O
trained	O
only	O
through	O
self	O
play	O
,	O
that	O
had	O
surpassed	O
all	O
previous	O
versions	O
,	O
beating	O
the	O
Ke	O
Jie	O
version	O
in	O
89	O
out	O
of	O
100	O
games	O
.	O
</s>
<s>
After	O
the	O
basic	O
principles	O
of	O
AlphaGo	B-Application
were	O
published	O
in	O
the	O
journal	O
Nature	O
,	O
other	O
teams	O
have	O
been	O
able	O
to	O
produce	O
high-level	O
programs	O
.	O
</s>
<s>
By	O
2017	O
,	O
both	O
Zen	B-Application
and	O
Tencent	O
's	O
project	O
Fine	B-Application
Art	I-Application
were	O
capable	O
of	O
defeating	O
very	O
high-level	O
professionals	O
some	O
of	O
the	O
time	O
.	O
</s>
<s>
The	O
open	B-Application
source	I-Application
Leela	B-Application
Zero	I-Application
engine	O
was	O
created	O
as	O
well	O
.	O
</s>
<s>
For	O
a	O
long	O
time	O
,	O
it	O
was	O
a	O
widely	O
held	O
opinion	O
that	O
computer	B-Application
Go	I-Application
posed	O
a	O
problem	O
fundamentally	O
different	O
from	O
computer	B-Application
chess	I-Application
.	O
</s>
<s>
Many	O
considered	O
a	O
strong	O
Go-playing	O
program	O
something	O
that	O
could	O
be	O
achieved	O
only	O
in	O
the	O
far	O
future	O
,	O
as	O
a	O
result	O
of	O
fundamental	O
advances	O
in	O
general	O
artificial	B-Application
intelligence	I-Application
technology	O
.	O
</s>
<s>
Therefore	O
,	O
a	O
large	O
part	O
of	O
the	O
computer	B-Application
Go	I-Application
development	O
effort	O
was	O
during	O
these	O
times	O
focused	O
on	O
ways	O
of	O
representing	O
human-like	O
expert	O
knowledge	O
and	O
combining	O
this	O
with	O
local	O
search	O
to	O
answer	O
questions	O
of	O
a	O
tactical	O
nature	O
.	O
</s>
<s>
The	O
large	O
board	O
size	O
prevents	O
an	O
alpha-beta	B-Algorithm
searcher	I-Algorithm
from	O
achieving	O
deep	O
look-ahead	O
without	O
significant	O
search	O
extensions	O
or	O
pruning	B-Algorithm
heuristics	B-Algorithm
.	O
</s>
<s>
In	O
2002	O
,	O
a	O
computer	B-Application
program	I-Application
called	O
MIGOS	O
(	O
MIni	O
GO	O
Solver	O
)	O
completely	O
solved	O
the	O
game	O
of	O
Go	O
for	O
the	O
5×5	O
board	O
.	O
</s>
<s>
Continuing	O
the	O
comparison	O
to	O
chess	B-Application
,	O
Go	O
moves	O
are	O
not	O
as	O
limited	O
by	O
the	O
rules	O
of	O
the	O
game	O
.	O
</s>
<s>
For	O
the	O
first	O
move	O
in	O
chess	B-Application
,	O
the	O
player	O
has	O
twenty	O
choices	O
.	O
</s>
<s>
In	O
chess	B-Application
,	O
many	O
future	O
positions	O
in	O
a	O
tree	O
are	O
direct	O
wins	O
for	O
one	O
side	O
,	O
and	O
boards	O
have	O
a	O
reasonable	O
heuristic	B-Algorithm
for	O
evaluation	O
in	O
simple	O
material	O
counting	O
,	O
as	O
well	O
as	O
certain	O
positional	O
factors	O
such	O
as	O
pawn	O
structure	O
.	O
</s>
<s>
The	O
value	O
of	O
a	O
Go	O
position	O
depends	O
on	O
a	O
complex	O
analysis	O
to	O
determine	O
whether	O
or	O
not	O
the	O
group	O
is	O
alive	O
,	O
which	O
stones	O
can	O
be	O
connected	O
to	O
one	O
another	O
,	O
and	O
heuristics	B-Algorithm
around	O
the	O
extent	O
to	O
which	O
a	O
strong	O
position	O
has	O
influence	O
,	O
or	O
the	O
extent	O
to	O
which	O
a	O
weak	O
position	O
can	O
be	O
attacked	O
.	O
</s>
<s>
Poor	O
evaluation	O
of	O
board	O
states	O
will	O
cause	O
the	O
AI	B-Application
to	O
work	O
toward	O
positions	O
it	O
incorrectly	O
believes	O
favor	O
it	O
,	O
but	O
actually	O
do	O
not	O
.	O
</s>
<s>
This	O
general	O
class	O
of	O
problems	O
is	O
known	O
as	O
life	B-Device
and	I-Device
death	I-Device
.	O
</s>
<s>
Knowledge-based	O
AI	B-Application
systems	O
sometimes	O
attempted	O
to	O
understand	O
the	O
life	B-Device
and	I-Device
death	I-Device
status	O
of	O
groups	O
on	O
the	O
board	O
.	O
</s>
<s>
The	O
most	O
direct	O
approach	O
is	O
to	O
perform	O
a	O
tree	B-Algorithm
search	I-Algorithm
on	O
the	O
moves	O
which	O
potentially	O
affect	O
the	O
stones	O
in	O
question	O
,	O
and	O
then	O
to	O
record	O
the	O
status	O
of	O
the	O
stones	O
at	O
the	O
end	O
of	O
the	O
main	O
line	O
of	O
play	O
.	O
</s>
<s>
This	O
implies	O
that	O
some	O
heuristic	B-Algorithm
must	O
be	O
applied	O
to	O
select	O
which	O
moves	O
to	O
consider	O
.	O
</s>
<s>
The	O
net	O
effect	O
is	O
that	O
for	O
any	O
given	O
program	O
,	O
there	O
is	O
a	O
trade-off	O
between	O
playing	O
speed	O
and	O
life	B-Device
and	I-Device
death	I-Device
reading	O
abilities	O
.	O
</s>
<s>
An	O
issue	O
that	O
all	O
Go	B-Application
programs	I-Application
must	O
tackle	O
is	O
how	O
to	O
represent	O
the	O
current	O
state	O
of	O
the	O
game	O
.	O
</s>
<s>
In	O
general	O
,	O
machine	O
learning	O
programs	O
stop	O
there	O
at	O
this	O
simplest	O
form	O
and	O
let	O
the	O
organic	O
AIs	B-Application
come	O
to	O
their	O
own	O
understanding	O
of	O
the	O
meaning	O
of	O
the	O
board	O
,	O
likely	O
simply	O
using	O
Monte	B-Algorithm
Carlo	I-Algorithm
playouts	O
to	O
"	O
score	O
"	O
a	O
board	O
as	O
good	O
or	O
bad	O
for	O
a	O
player	O
.	O
</s>
<s>
"	O
Classic	O
"	O
AI	B-Application
programs	O
that	O
attempted	O
to	O
directly	O
model	O
a	O
human	O
's	O
strategy	O
might	O
go	O
further	O
,	O
however	O
,	O
such	O
as	O
layering	O
on	O
data	O
such	O
as	O
stones	O
believed	O
to	O
be	O
dead	O
,	O
stones	O
that	O
are	O
unconditionally	O
alive	O
,	O
stones	O
in	O
a	O
seki	O
state	O
of	O
mutual	O
life	O
,	O
and	O
so	O
forth	O
in	O
their	O
representation	O
of	O
the	O
state	O
of	O
the	O
game	O
.	O
</s>
<s>
Historically	O
,	O
symbolic	B-General_Concept
artificial	I-General_Concept
intelligence	I-General_Concept
techniques	O
have	O
been	O
used	O
to	O
approach	O
the	O
problem	O
of	O
Go	O
AI	B-Application
.	O
</s>
<s>
Neural	B-Architecture
networks	I-Architecture
began	O
to	O
be	O
tried	O
as	O
an	O
alternative	O
approach	O
in	O
the	O
2000s	O
decade	O
,	O
as	O
they	O
required	O
immense	O
computing	O
power	O
that	O
was	O
expensive-to-impossible	O
to	O
reach	O
in	O
earlier	O
decades	O
.	O
</s>
<s>
These	O
approaches	O
attempt	O
to	O
mitigate	O
the	O
problems	O
of	O
the	O
game	O
of	O
Go	O
having	O
a	O
high	O
branching	B-Data_Structure
factor	I-Data_Structure
and	O
numerous	O
other	O
difficulties	O
.	O
</s>
<s>
some	O
form	O
of	O
tree	B-Algorithm
search	I-Algorithm
,	O
</s>
<s>
pattern	B-Language
matching	I-Language
and	O
knowledge-based	B-General_Concept
systems	I-General_Concept
,	O
</s>
<s>
the	O
application	O
of	O
Monte	B-Algorithm
Carlo	I-Algorithm
methods	I-Algorithm
,	O
</s>
<s>
One	O
traditional	B-General_Concept
AI	I-General_Concept
technique	O
for	O
creating	O
game	O
playing	O
software	O
is	O
to	O
use	O
a	O
minimax	B-Algorithm
tree	B-Algorithm
search	I-Algorithm
.	O
</s>
<s>
This	O
involves	O
playing	O
out	O
all	O
hypothetical	O
moves	O
on	O
the	O
board	O
up	O
to	O
a	O
certain	O
point	O
,	O
then	O
using	O
an	O
evaluation	B-General_Concept
function	I-General_Concept
to	O
estimate	O
the	O
value	O
of	O
that	O
position	O
for	O
the	O
current	O
player	O
.	O
</s>
<s>
While	O
tree	B-Algorithm
searches	I-Algorithm
have	O
been	O
very	O
effective	O
in	O
computer	B-Application
chess	I-Application
,	O
they	O
have	O
seen	O
less	O
success	O
in	O
Computer	B-Application
Go	I-Application
programs	O
.	O
</s>
<s>
This	O
is	O
partly	O
because	O
it	O
has	O
traditionally	O
been	O
difficult	O
to	O
create	O
an	O
effective	O
evaluation	B-General_Concept
function	I-General_Concept
for	O
a	O
Go	O
board	O
,	O
and	O
partly	O
because	O
the	O
large	O
number	O
of	O
possible	O
moves	O
each	O
side	O
can	O
make	O
each	O
leads	O
to	O
a	O
high	O
branching	B-Data_Structure
factor	I-Data_Structure
.	O
</s>
<s>
Pruning	B-Algorithm
techniques	O
such	O
as	O
alpha	B-Algorithm
–	I-Algorithm
beta	I-Algorithm
pruning	I-Algorithm
,	O
Principal	B-General_Concept
Variation	I-General_Concept
Search	I-General_Concept
,	O
and	O
MTD(f )	O
can	O
reduce	O
the	O
effective	O
branching	B-Data_Structure
factor	I-Data_Structure
without	O
loss	O
of	O
strength	O
.	O
</s>
<s>
In	O
tactical	O
areas	O
such	O
as	O
life	B-Device
and	I-Device
death	I-Device
,	O
Go	O
is	O
particularly	O
amenable	O
to	O
caching	O
techniques	O
such	O
as	O
transposition	B-General_Concept
tables	I-General_Concept
.	O
</s>
<s>
These	O
can	O
reduce	O
the	O
amount	O
of	O
repeated	O
effort	O
,	O
especially	O
when	O
combined	O
with	O
an	O
iterative	B-Algorithm
deepening	I-Algorithm
approach	O
.	O
</s>
<s>
In	O
order	O
to	O
quickly	O
store	O
a	O
full-sized	O
Go	O
board	O
in	O
a	O
transposition	B-General_Concept
table	I-General_Concept
,	O
a	O
hashing	B-Error_Name
technique	O
for	O
mathematically	O
summarizing	O
is	O
generally	O
necessary	O
.	O
</s>
<s>
Zobrist	B-Application
hashing	I-Application
is	O
very	O
popular	O
in	O
Go	B-Application
programs	I-Application
because	O
it	O
has	O
low	O
collision	O
rates	O
,	O
and	O
can	O
be	O
iteratively	O
updated	O
at	O
each	O
move	O
with	O
just	O
two	O
XORs	B-Application
,	O
rather	O
than	O
being	O
calculated	O
from	O
scratch	O
.	O
</s>
<s>
Even	O
using	O
these	O
performance-enhancing	O
techniques	O
,	O
full	O
tree	B-Algorithm
searches	I-Algorithm
on	O
a	O
full-sized	O
board	O
are	O
still	O
prohibitively	O
slow	O
.	O
</s>
<s>
Searches	O
can	O
be	O
sped	O
up	O
by	O
using	O
large	O
amounts	O
of	O
domain	O
specific	O
pruning	B-Algorithm
techniques	O
,	O
such	O
as	O
not	O
considering	O
moves	O
where	O
your	O
opponent	O
is	O
already	O
strong	O
,	O
and	O
selective	O
extensions	O
like	O
always	O
considering	O
moves	O
next	O
to	O
groups	O
of	O
stones	O
which	O
are	O
about	O
to	O
be	O
captured	O
.	O
</s>
<s>
Results	O
of	O
computer	O
competitions	O
show	O
that	O
pattern	B-Language
matching	I-Language
techniques	O
for	O
choosing	O
a	O
handful	O
of	O
appropriate	O
moves	O
combined	O
with	O
fast	O
localized	O
tactical	O
searches	O
(	O
explained	O
above	O
)	O
were	O
once	O
sufficient	O
to	O
produce	O
a	O
competitive	O
program	O
.	O
</s>
<s>
For	O
example	O
,	O
GNU	B-Application
Go	I-Application
was	O
competitive	O
until	O
2008	O
.	O
</s>
<s>
AI	B-Application
work	O
in	O
the	O
1990s	O
often	O
involved	O
attempting	O
to	O
"	O
teach	O
"	O
the	O
AI	B-Application
human-style	O
heuristics	B-Algorithm
of	O
Go	O
knowledge	O
.	O
</s>
<s>
In	O
1996	O
,	O
Tim	O
Klinger	O
and	O
David	O
Mechner	O
acknowledged	O
the	O
beginner-level	O
strength	O
of	O
the	O
best	O
AIs	B-Application
and	O
argued	O
that	O
"	O
it	O
is	O
our	O
belief	O
that	O
with	O
better	O
tools	O
for	O
representing	O
and	O
maintaining	O
Go	O
knowledge	O
,	O
it	O
will	O
be	O
possible	O
to	O
develop	O
stronger	O
Go	O
programs.	O
"	O
</s>
<s>
In	O
2001	O
,	O
one	O
paper	O
concluded	O
that	O
"	O
Go	B-Application
programs	I-Application
are	O
still	O
lacking	O
in	O
both	O
quality	O
and	O
quantity	O
of	O
knowledge	O
,	O
"	O
and	O
that	O
fixing	O
this	O
would	O
improve	O
Go	O
AI	B-Application
performance	O
.	O
</s>
<s>
The	O
programmer	O
's	O
task	O
is	O
to	O
take	O
these	O
heuristics	B-Algorithm
,	O
formalize	O
them	O
into	O
computer	B-Application
code	I-Application
,	O
and	O
utilize	O
pattern	B-Language
matching	I-Language
and	O
pattern	O
recognition	O
algorithms	O
to	O
recognize	O
when	O
these	O
rules	O
apply	O
.	O
</s>
<s>
It	O
is	O
also	O
important	O
to	O
be	O
able	O
to	O
"	O
score	O
"	O
these	O
heuristics	B-Algorithm
so	O
that	O
when	O
they	O
offer	O
conflicting	O
advice	O
,	O
the	O
system	O
has	O
ways	O
to	O
determine	O
which	O
heuristic	B-Algorithm
is	O
more	O
important	O
and	O
applicable	O
to	O
the	O
situation	O
.	O
</s>
<s>
One	O
major	O
alternative	O
to	O
using	O
hand-coded	O
knowledge	O
and	O
searches	O
is	O
the	O
use	O
of	O
Monte	B-Algorithm
Carlo	I-Algorithm
methods	I-Algorithm
.	O
</s>
<s>
No	O
potentially	O
fallible	O
knowledge-based	B-General_Concept
system	I-General_Concept
is	O
required	O
.	O
</s>
<s>
Some	O
programs	O
which	O
use	O
Monte-Carlo	O
techniques	O
are	O
Fuego	O
,	O
The	O
Many	O
Faces	O
of	O
Go	O
v12	O
,	O
Leela	B-Application
,	O
MoGo	O
,	O
Crazy	B-Application
Stone	I-Application
,	O
MyGoFriend	B-Application
,	O
and	O
Zen	B-Application
.	O
</s>
<s>
In	O
2006	O
,	O
a	O
new	O
search	O
technique	O
,	O
upper	O
confidence	O
bounds	O
applied	O
to	O
trees	O
(	O
UCT	O
)	O
,	O
was	O
developed	O
and	O
applied	O
to	O
many	O
9x9	O
Monte-Carlo	O
Go	B-Application
programs	I-Application
with	O
excellent	O
results	O
.	O
</s>
<s>
Successful	O
early	O
applications	O
of	O
UCT	O
methods	O
to	O
19x19	O
Go	O
include	O
MoGo	O
,	O
Crazy	B-Application
Stone	I-Application
,	O
and	O
Mango	O
.	O
</s>
<s>
That	O
said	O
,	O
this	O
tendency	O
has	O
persisted	O
in	O
AlphaGo	B-Application
's	O
playstyle	O
with	O
dominant	O
results	O
,	O
so	O
this	O
may	O
be	O
more	O
of	O
a	O
"	O
quirk	O
"	O
than	O
a	O
"	O
weakness.	O
"	O
</s>
<s>
The	O
skill	O
level	O
of	O
knowledge-based	B-General_Concept
systems	I-General_Concept
is	O
closely	O
linked	O
to	O
the	O
knowledge	O
of	O
their	O
programmers	O
and	O
associated	O
domain	O
experts	O
.	O
</s>
<s>
This	O
limitation	O
has	O
made	O
it	O
difficult	O
to	O
program	O
truly	O
strong	O
AIs	B-Application
.	O
</s>
<s>
The	O
software	O
will	O
then	O
automatically	O
generates	O
its	O
own	O
sense	O
of	O
patterns	O
,	O
heuristics	B-Algorithm
,	O
and	O
strategies	O
,	O
in	O
theory	O
.	O
</s>
<s>
This	O
is	O
generally	O
done	O
by	O
allowing	O
a	O
neural	B-Architecture
network	I-Architecture
or	O
genetic	B-Algorithm
algorithm	I-Algorithm
to	O
either	O
review	O
a	O
large	O
database	O
of	O
professional	O
games	O
,	O
or	O
play	O
many	O
games	O
against	O
itself	O
or	O
other	O
people	O
or	O
programs	O
.	O
</s>
<s>
For	O
example	O
,	O
Crazy	B-Application
Stone	I-Application
learns	O
move	O
generation	O
patterns	O
from	O
several	O
hundred	O
sample	O
games	O
,	O
using	O
a	O
generalization	O
of	O
the	O
Elo	O
rating	O
system	O
.	O
</s>
<s>
The	O
most	O
famous	O
example	O
of	O
this	O
approach	O
is	O
AlphaGo	B-Application
,	O
which	O
proved	O
far	O
more	O
effective	O
than	O
previous	O
AIs	B-Application
.	O
</s>
<s>
AlphaGo	B-Application
used	O
Monte	B-Application
Carlo	I-Application
tree	I-Application
search	I-Application
to	O
score	O
the	O
resulting	O
positions	O
.	O
</s>
<s>
A	O
later	O
version	O
of	O
AlphaGo	B-Application
,	O
AlphaGoZero	O
,	O
eschewed	O
learning	O
from	O
existing	O
Go	O
games	O
,	O
and	O
instead	O
learnt	O
only	O
from	O
playing	O
itself	O
repeatedly	O
.	O
</s>
<s>
Other	O
earlier	O
programs	O
using	O
neural	B-Architecture
nets	I-Architecture
include	O
NeuroGo	O
and	O
WinHonte	O
.	O
</s>
<s>
Computer	B-Application
Go	I-Application
research	O
results	O
are	O
being	O
applied	O
to	O
other	O
similar	O
fields	O
such	O
as	O
cognitive	O
science	O
,	O
pattern	O
recognition	O
and	O
machine	O
learning	O
.	O
</s>
<s>
Combinatorial	O
Game	O
Theory	O
,	O
a	O
branch	O
of	O
applied	O
mathematics	O
,	O
is	O
a	O
topic	O
relevant	O
to	O
computer	B-Application
Go	I-Application
.	O
</s>
<s>
(	O
In	O
practice	O
,	O
strong	O
Monte	B-Algorithm
Carlo	I-Algorithm
algorithms	O
can	O
still	O
handle	O
normal	O
Go	O
endgame	O
situations	O
well	O
enough	O
,	O
and	O
the	O
most	O
complicated	O
classes	O
of	O
life-and-death	O
endgame	O
problems	O
are	O
unlikely	O
to	O
come	O
up	O
in	O
a	O
high-level	O
game	O
.	O
)	O
</s>
<s>
Various	O
difficult	O
combinatorial	O
problems	O
(	O
any	O
NP-hard	O
problem	O
)	O
can	O
be	O
converted	O
to	O
Go-like	O
problems	O
on	O
a	O
sufficiently	O
large	O
board	O
;	O
however	O
,	O
the	O
same	O
is	O
true	O
for	O
other	O
abstract	O
board	O
games	O
,	O
including	O
chess	B-Application
and	O
minesweeper	B-Application
,	O
when	O
suitably	O
generalized	O
to	O
a	O
board	O
of	O
arbitrary	O
size	O
.	O
</s>
<s>
NP-complete	O
problems	O
do	O
not	O
tend	O
in	O
their	O
general	O
case	O
to	O
be	O
easier	O
for	O
unaided	O
humans	O
than	O
for	O
suitably	O
programmed	O
computers	O
:	O
unaided	O
humans	O
are	O
much	O
worse	O
than	O
computers	O
at	O
solving	O
,	O
for	O
example	O
,	O
instances	O
of	O
the	O
subset	B-Algorithm
sum	I-Algorithm
problem	I-Algorithm
.	O
</s>
<s>
Zen	B-Application
,	O
by	O
Yoji	O
Ojima	O
aka	O
Yamato	O
(	O
sold	O
as	O
Tencho	B-Application
no	I-Application
Igo	I-Application
in	O
Japan	O
)	O
;	O
parallel	O
version	O
by	O
Hideki	O
Kato	O
.	O
</s>
<s>
AlphaGo	B-Application
,	O
developed	O
by	O
Google	B-Application
DeepMind	I-Application
,	O
made	O
a	O
significant	O
advance	O
by	O
beating	O
a	O
professional	O
human	O
player	O
in	O
October	O
2015	O
,	O
using	O
techniques	O
that	O
combined	O
deep	B-Algorithm
learning	I-Algorithm
and	O
Monte	B-Application
Carlo	I-Application
tree	I-Application
search	I-Application
.	O
</s>
<s>
AlphaGo	B-Application
was	O
significantly	O
more	O
powerful	O
than	O
previous	O
Go	B-Application
programs	I-Application
,	O
and	O
was	O
the	O
first	O
to	O
beat	O
a	O
9	O
dan	O
human	O
professional	O
in	O
a	O
game	O
without	O
handicaps	O
on	O
a	O
full-sized	O
board	O
.	O
</s>
<s>
Work	O
on	O
Go	O
AI	B-Application
since	O
has	O
largely	O
consisted	O
of	O
emulating	O
the	O
techniques	O
used	O
to	O
build	O
AlphaGo	B-Application
,	O
which	O
proved	O
so	O
much	O
stronger	O
than	O
everything	O
else	O
.	O
</s>
<s>
Several	O
annual	O
competitions	O
take	O
place	O
between	O
Go	O
computer	B-Application
programs	I-Application
,	O
including	O
Go	O
events	O
at	O
the	O
Computer	O
Olympiad	O
.	O
</s>
<s>
Regular	O
,	O
less	O
formal	O
,	O
competitions	O
between	O
programs	O
used	O
to	O
occur	O
on	O
the	O
KGS	B-Application
Go	I-Application
Server	I-Application
(	O
monthly	O
)	O
and	O
the	O
Computer	B-Application
Go	I-Application
Server	O
(	O
continuous	O
)	O
.	O
</s>
<s>
Many	O
programs	O
are	O
available	O
that	O
allow	O
computer	B-Application
Go	I-Application
engines	O
to	O
play	O
against	O
each	O
other	O
;	O
they	O
almost	O
always	O
communicate	O
via	O
the	O
Go	B-Application
Text	I-Application
Protocol	I-Application
(	O
GTP	O
)	O
.	O
</s>
<s>
The	O
first	O
computer	B-Application
Go	I-Application
competition	O
was	O
sponsored	O
by	O
Acornsoft	O
,	O
and	O
the	O
first	O
regular	O
ones	O
by	O
USENIX	B-Operating_System
.	O
</s>
<s>
One	O
of	O
the	O
early	O
drivers	O
of	O
computer	B-Application
Go	I-Application
research	O
was	O
the	O
Ing	O
Prize	O
,	O
a	O
relatively	O
large	O
money	O
award	O
sponsored	O
by	O
Taiwanese	O
banker	O
Ing	O
Chang-ki	O
,	O
offered	O
annually	O
between	O
1985	O
and	O
2000	O
at	O
the	O
World	O
Computer	B-Application
Go	I-Application
Congress	O
(	O
or	O
Ing	O
Cup	O
)	O
.	O
</s>
<s>
Many	O
other	O
large	O
regional	O
Go	O
tournaments	O
(	O
"	O
congresses	O
"	O
)	O
had	O
an	O
attached	O
computer	B-Application
Go	I-Application
event	O
.	O
</s>
<s>
The	O
European	O
Go	O
Congress	O
has	O
sponsored	O
a	O
computer	O
tournament	O
since	O
1987	O
,	O
and	O
the	O
USENIX	B-Operating_System
event	O
evolved	O
into	O
the	O
US/North	O
American	O
Computer	B-Application
Go	I-Application
Championship	O
,	O
held	O
annually	O
from	O
1988	O
–	O
2000	O
at	O
the	O
US	O
Go	O
Congress	O
.	O
</s>
<s>
Japan	O
started	O
sponsoring	O
computer	B-Application
Go	I-Application
competitions	O
in	O
1995	O
.	O
</s>
<s>
The	O
Computer	B-Application
Go	I-Application
UEC	O
Cup	O
has	O
been	O
held	O
annually	O
since	O
2007	O
.	O
</s>
<s>
The	O
main	O
problem	O
is	O
that	O
Go	O
playing	O
software	O
,	O
which	O
usually	O
communicates	O
using	O
the	O
standardized	O
Go	B-Application
Text	I-Application
Protocol	I-Application
(	O
GTP	O
)	O
,	O
will	O
not	O
always	O
agree	O
with	O
respect	O
to	O
the	O
alive	O
or	O
dead	O
status	O
of	O
stones	O
.	O
</s>
<s>
In	O
practice	O
,	O
such	O
as	O
on	O
the	O
KGS	B-Application
Go	I-Application
Server	I-Application
,	O
the	O
server	O
can	O
mediate	O
a	O
dispute	O
by	O
sending	O
a	O
special	O
GTP	O
command	O
to	O
the	O
two	O
client	O
programs	O
indicating	O
they	O
should	O
continue	O
placing	O
stones	O
until	O
there	O
is	O
no	O
question	O
about	O
the	O
status	O
of	O
any	O
particular	O
group	O
(	O
all	O
dead	O
stones	O
have	O
been	O
captured	O
)	O
.	O
</s>
<s>
Nevertheless	O
,	O
most	O
modern	O
Go	B-Application
Programs	I-Application
support	O
Japanese	O
rules	O
against	O
humans	O
.	O
</s>
