<s>
Darkforest	B-Application
is	O
a	O
computer	B-Application
go	I-Application
program	O
developed	O
by	O
Meta	O
Platforms	O
,	O
based	O
on	O
deep	B-Algorithm
learning	I-Algorithm
techniques	O
using	O
a	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
.	O
</s>
<s>
Its	O
updated	O
version	O
Darkfores2	B-Application
combines	O
the	O
techniques	O
of	O
its	O
predecessor	O
with	O
Monte	B-Application
Carlo	I-Application
tree	I-Application
search	I-Application
.	O
</s>
<s>
Darkforest	B-Application
is	O
of	O
similar	O
strength	O
to	O
programs	O
like	O
CrazyStone	B-Application
and	O
Zen	O
.	O
</s>
<s>
Google	B-Application
's	I-Application
AlphaGo	B-Application
program	O
won	O
against	O
a	O
professional	O
player	O
in	O
October	O
2015	O
using	O
a	O
similar	O
combination	O
of	O
techniques	O
.	O
</s>
<s>
Darkforest	B-Application
is	O
named	O
after	O
Liu	O
Cixin	O
's	O
science	O
fiction	O
novel	O
The	O
Dark	O
Forest	O
.	O
</s>
<s>
Go	O
’s	O
high	O
branching	B-Data_Structure
factor	I-Data_Structure
makes	O
traditional	O
search	O
techniques	O
ineffective	O
,	O
even	O
on	O
cutting-edge	O
hardware	O
,	O
and	O
Go	O
’s	O
evaluation	B-General_Concept
function	I-General_Concept
could	O
change	O
drastically	O
with	O
one	O
stone	O
change	O
.	O
</s>
<s>
However	O
,	O
by	O
using	O
a	O
Deep	B-Architecture
Convolutional	I-Architecture
Neural	I-Architecture
Network	I-Architecture
designed	O
for	O
long-term	O
predictions	O
,	O
Darkforest	B-Application
has	O
been	O
able	O
to	O
substantially	O
improve	O
the	O
win	O
rate	O
for	O
bots	O
over	O
more	O
traditional	O
Monte	B-Application
Carlo	I-Application
Tree	I-Application
Search	I-Application
based	O
approaches	O
.	O
</s>
<s>
Against	O
human	O
players	O
,	O
Darkfores2	B-Application
achieves	O
a	O
stable	O
3d	O
ranking	O
on	O
KGS	B-Application
Go	I-Application
Server	I-Application
,	O
which	O
roughly	O
corresponds	O
to	O
an	O
advanced	O
amateur	O
human	O
player	O
.	O
</s>
<s>
However	O
,	O
after	O
adding	O
Monte	B-Application
Carlo	I-Application
Tree	I-Application
Search	I-Application
to	O
Darkfores2	B-Application
to	O
create	O
a	O
much	O
stronger	O
player	O
named	O
darkfmcts3	O
,	O
it	O
can	O
achieve	O
a	O
5d	O
ranking	O
on	O
the	O
KGS	B-Application
Go	I-Application
Server	I-Application
.	O
</s>
<s>
darkfmcts3	O
is	O
on	O
par	O
with	O
state-of-the-art	O
Go	O
AIs	O
such	O
as	O
Zen	O
,	O
DolBaram	O
and	O
Crazy	B-Application
Stone	I-Application
but	O
lags	O
behind	O
AlphaGo	B-Application
.	O
</s>
<s>
After	O
Google	B-Application
's	I-Application
AlphaGo	B-Application
won	O
against	O
Fan	O
Hui	O
in	O
2015	O
,	O
Facebook	O
made	O
its	O
AI	O
's	O
hardware	O
designs	O
public	O
,	O
alongside	O
releasing	O
the	O
code	O
behind	O
DarkForest	B-Application
as	O
open-source	O
,	O
along	O
with	O
heavy	O
recruiting	O
to	O
strengthen	O
its	O
team	O
of	O
AI	O
engineers	O
.	O
</s>
<s>
Darkforest	B-Application
uses	O
a	O
neural	O
network	O
to	O
sort	O
through	O
the	O
10100	O
board	O
positions	O
,	O
and	O
find	O
the	O
most	O
powerful	O
next	O
move	O
.	O
</s>
<s>
However	O
,	O
neural	O
networks	O
alone	O
cannot	O
match	O
the	O
level	O
of	O
good	O
amateur	O
players	O
or	O
the	O
best	O
search-based	O
Go	O
engines	O
,	O
and	O
so	O
Darkfores2	B-Application
combines	O
the	O
neural	O
network	O
approach	O
with	O
a	O
search-based	O
machine	O
.	O
</s>
<s>
A	O
database	O
of	O
250,000	O
real	O
Go	O
games	O
were	O
used	O
in	O
the	O
development	O
of	O
Darkforest	B-Application
,	O
with	O
220,000	O
used	O
as	O
a	O
training	O
set	O
and	O
the	O
rest	O
used	O
to	O
test	O
the	O
neural	O
network	O
's	O
ability	O
to	O
predict	O
the	O
next	O
moves	O
played	O
in	O
the	O
real	O
games	O
.	O
</s>
<s>
This	O
allows	O
Darkforest	B-Application
to	O
accurately	O
evaluate	O
the	O
global	O
state	O
of	O
the	O
board	O
,	O
but	O
local	O
tactics	O
were	O
still	O
poor	O
.	O
</s>
<s>
Combining	O
these	O
two	O
approaches	O
is	O
difficult	O
because	O
search-based	O
engines	O
work	O
much	O
faster	O
than	O
neural	O
networks	O
,	O
a	O
problem	O
which	O
was	O
solved	O
in	O
Darkfores2	B-Application
by	O
running	O
the	O
processes	O
in	O
parallel	O
with	O
frequent	O
communication	O
between	O
the	O
two	O
.	O
</s>
<s>
It	O
has	O
been	O
noted	O
that	O
Darkforest	B-Application
still	O
has	O
flaws	O
in	O
its	O
play	O
style	O
.	O
</s>
<s>
The	O
family	O
of	O
Darkforest	B-Application
computer	B-Application
go	I-Application
programs	O
is	O
based	O
on	O
convolution	B-Architecture
neural	I-Architecture
networks	I-Architecture
.	O
</s>
<s>
The	O
most	O
recent	O
advances	O
in	O
Darkfmcts3	O
combined	O
convolutional	B-Architecture
neural	I-Architecture
networks	I-Architecture
with	O
more	O
traditional	O
Monte	B-Application
Carlo	I-Application
tree	I-Application
search	I-Application
.	O
</s>
<s>
Darkfmcts3	O
is	O
the	O
most	O
advanced	O
version	O
of	O
Darkforest	B-Application
,	O
which	O
combines	O
Facebook	O
's	O
most	O
advanced	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
architecture	O
from	O
Darkfores2	B-Application
with	O
a	O
Monte	B-Application
Carlo	I-Application
tree	I-Application
search	I-Application
.	O
</s>
<s>
Darkfmcts3	O
relies	O
on	O
a	O
convolution	B-Architecture
neural	I-Architecture
networks	I-Architecture
that	O
predicts	O
the	O
next	O
k	O
moves	O
based	O
on	O
the	O
current	O
state	O
of	O
play	O
.	O
</s>
<s>
Each	O
convolutional	O
layer	O
is	O
followed	O
by	O
a	O
rectified	B-Algorithm
linear	I-Algorithm
unit	I-Algorithm
,	O
a	O
popular	O
activation	O
function	O
for	O
deep	O
neural	O
networks	O
.	O
</s>
<s>
A	O
key	O
innovation	O
of	O
Darkfmct3	O
compared	O
to	O
previous	O
approaches	O
is	O
that	O
it	O
uses	O
only	O
one	O
softmax	B-Algorithm
function	I-Algorithm
to	O
predict	O
the	O
next	O
move	O
,	O
which	O
enables	O
the	O
approach	O
to	O
reduce	O
the	O
overall	O
number	O
of	O
parameters	O
.	O
</s>
<s>
The	O
learning	O
rate	O
was	O
determined	O
by	O
vanilla	O
stochastic	B-Algorithm
gradient	I-Algorithm
descent	I-Algorithm
.	O
</s>
<s>
Darkfmct3	O
synchronously	O
couples	O
a	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
with	O
a	O
Monte	B-Application
Carlo	I-Application
tree	I-Application
search	I-Application
.	O
</s>
<s>
Because	O
the	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
is	O
computationally	O
taxing	O
,	O
the	O
Monte	B-Application
Carlo	I-Application
tree	I-Application
search	I-Application
focuses	O
computation	O
on	O
the	O
more	O
likely	O
game	O
play	O
trajectories	O
.	O
</s>
<s>
By	O
running	O
the	O
neural	O
network	O
synchronously	O
with	O
the	O
Monte	B-Application
Carlo	I-Application
tree	I-Application
search	I-Application
,	O
it	O
is	O
possible	O
to	O
guarantee	O
that	O
each	O
node	O
is	O
expanded	O
by	O
the	O
moves	O
predicted	O
by	O
the	O
neural	O
network	O
.	O
</s>
<s>
Darkfores2	B-Application
beats	O
Darkforest	B-Application
,	O
its	O
neural	O
network-only	O
predecessor	O
,	O
around	O
90%	O
of	O
the	O
time	O
,	O
and	O
Pachi	O
,	O
one	O
of	O
the	O
best	O
search-based	O
engines	O
,	O
around	O
95%	O
of	O
the	O
time	O
.	O
</s>
<s>
On	O
the	O
Kyu	O
rating	O
system	O
,	O
Darkforest	B-Application
holds	O
a	O
1-2d	O
level	O
.	O
</s>
<s>
Darkfores2	B-Application
achieves	O
a	O
stable	O
3d	O
level	O
on	O
KGS	B-Application
Go	I-Application
Server	I-Application
as	O
a	O
ranked	O
bot	O
.	O
</s>
<s>
With	O
the	O
added	O
Monte	B-Application
Carlo	I-Application
tree	I-Application
search	I-Application
,	O
Darkfmcts3	O
with	O
5,000	O
rollouts	O
beats	O
Pachi	O
with	O
10k	O
rollouts	O
in	O
all	O
250	O
games	O
;	O
with	O
75k	O
rollouts	O
it	O
achieves	O
a	O
stable	O
5d	O
level	O
in	O
KGS	O
server	O
,	O
on	O
par	O
with	O
state-of-the-art	O
Go	O
AIs	O
(	O
e.g.	O
,	O
Zen	O
,	O
DolBaram	O
,	O
CrazyStone	B-Application
)	O
;	O
with	O
110k	O
rollouts	O
,	O
it	O
won	O
the	O
3rd	O
place	O
in	O
January	O
KGS	O
Go	O
Tournament	O
.	O
</s>
