<s>
AlphaGo	B-Application
Zero	I-Application
is	O
a	O
version	O
of	O
DeepMind	B-Application
's	O
Go	B-Application
software	I-Application
AlphaGo	B-Application
.	O
</s>
<s>
AlphaGo	B-Application
's	O
team	O
published	O
an	O
article	O
in	O
the	O
journal	O
Nature	O
on	O
19	O
October	O
2017	O
,	O
introducing	O
AlphaGo	B-Application
Zero	I-Application
,	O
a	O
version	O
created	O
without	O
using	O
data	O
from	O
human	O
games	O
,	O
and	O
stronger	O
than	O
any	O
previous	O
version	O
.	O
</s>
<s>
By	O
playing	O
games	O
against	O
itself	O
,	O
AlphaGo	B-Application
Zero	I-Application
surpassed	O
the	O
strength	O
of	O
AlphaGo	B-Application
Lee	I-Application
in	O
three	O
days	O
by	O
winning	O
100	O
games	O
to	O
0	O
,	O
reached	O
the	O
level	O
of	O
AlphaGo	B-Application
Master	I-Application
in	O
21	O
days	O
,	O
and	O
exceeded	O
all	O
the	O
old	O
versions	O
in	O
40	O
days	O
.	O
</s>
<s>
Training	O
artificial	B-Application
intelligence	I-Application
(	O
AI	B-Application
)	O
without	O
datasets	O
derived	O
from	O
human	O
experts	O
has	O
significant	O
implications	O
for	O
the	O
development	O
of	O
AI	B-Application
with	O
superhuman	O
skills	O
because	O
expert	O
data	O
is	O
"	O
often	O
expensive	O
,	O
unreliable	O
or	O
simply	O
unavailable.	O
"	O
</s>
<s>
Demis	O
Hassabis	O
,	O
the	O
co-founder	O
and	O
CEO	O
of	O
DeepMind	B-Application
,	O
said	O
that	O
AlphaGo	B-Application
Zero	I-Application
was	O
so	O
powerful	O
because	O
it	O
was	O
"	O
no	O
longer	O
constrained	O
by	O
the	O
limits	O
of	O
human	O
knowledge	O
"	O
.	O
</s>
<s>
Furthermore	O
,	O
AlphaGo	B-Application
Zero	I-Application
performed	O
better	O
than	O
standard	O
reinforcement	O
deep	O
learning	O
models	O
(	O
such	O
as	O
DQN	O
implementations	O
)	O
due	O
to	O
its	O
integration	O
of	O
Monte	O
Carlo	O
tree	O
search	O
.	O
</s>
<s>
David	O
Silver	O
,	O
one	O
of	O
the	O
first	O
authors	O
of	O
DeepMind	B-Application
's	O
papers	O
published	O
in	O
Nature	O
on	O
AlphaGo	B-Application
,	O
said	O
that	O
it	O
is	O
possible	O
to	O
have	O
generalised	O
AI	B-Application
algorithms	O
by	O
removing	O
the	O
need	O
to	O
learn	O
from	O
humans	O
.	O
</s>
<s>
Google	O
later	O
developed	O
AlphaZero	B-Application
,	O
a	O
generalized	O
version	O
of	O
AlphaGo	B-Application
Zero	I-Application
that	O
could	O
play	O
chess	B-Application
and	O
Shōgi	B-Application
in	O
addition	O
to	O
Go	O
.	O
</s>
<s>
In	O
December	O
2017	O
,	O
AlphaZero	B-Application
beat	O
the	O
3-day	O
version	O
of	O
AlphaGo	B-Application
Zero	I-Application
by	O
winning	O
60	O
games	O
to	O
40	O
,	O
and	O
with	O
8	O
hours	O
of	O
training	O
it	O
outperformed	O
AlphaGo	B-Application
Lee	I-Application
on	O
an	O
Elo	O
scale	O
.	O
</s>
<s>
AlphaZero	B-Application
also	O
defeated	O
a	O
top	O
chess	B-Application
program	O
(	O
Stockfish	B-Application
)	O
and	O
a	O
top	O
Shōgi	B-Application
program	O
(	O
Elmo	B-Application
)	O
.	O
</s>
<s>
AlphaGo	B-Application
Zero	I-Application
's	O
neural	B-Architecture
network	I-Architecture
was	O
trained	O
using	O
TensorFlow	B-Language
,	O
with	O
64	O
GPU	B-Architecture
workers	O
and	O
19	O
CPU	O
parameter	O
servers	O
.	O
</s>
<s>
Only	O
four	O
TPUs	B-Device
were	O
used	O
for	O
inference	O
.	O
</s>
<s>
The	O
neural	B-Architecture
network	I-Architecture
initially	O
knew	O
nothing	O
about	O
Go	O
beyond	O
the	O
rules	O
.	O
</s>
<s>
Unlike	O
earlier	O
versions	O
of	O
AlphaGo	B-Application
,	O
Zero	O
only	O
perceived	O
the	O
board	O
's	O
stones	O
,	O
rather	O
than	O
having	O
some	O
rare	O
human-programmed	O
edge	O
cases	O
to	O
help	O
recognize	O
unusual	O
Go	O
board	O
positions	O
.	O
</s>
<s>
The	O
AI	B-Application
engaged	O
in	O
reinforcement	O
learning	O
,	O
playing	O
against	O
itself	O
until	O
it	O
could	O
anticipate	O
its	O
own	O
moves	O
and	O
how	O
those	O
moves	O
would	O
affect	O
the	O
game	O
's	O
outcome	O
.	O
</s>
<s>
In	O
the	O
first	O
three	O
days	O
AlphaGo	B-Application
Zero	I-Application
played	O
4.9	O
million	O
games	O
against	O
itself	O
in	O
quick	O
succession	O
.	O
</s>
<s>
It	O
appeared	O
to	O
develop	O
the	O
skills	O
required	O
to	O
beat	O
top	O
humans	O
within	O
just	O
a	O
few	O
days	O
,	O
whereas	O
the	O
earlier	O
AlphaGo	B-Application
took	O
months	O
of	O
training	O
to	O
achieve	O
the	O
same	O
level	O
.	O
</s>
<s>
For	O
comparison	O
,	O
the	O
researchers	O
also	O
trained	O
a	O
version	O
of	O
AlphaGo	B-Application
Zero	I-Application
using	O
human	O
games	O
,	O
AlphaGo	B-Application
Master	I-Application
,	O
and	O
found	O
that	O
it	O
learned	O
more	O
quickly	O
,	O
but	O
actually	O
performed	O
more	O
poorly	O
in	O
the	O
long	O
run	O
.	O
</s>
<s>
DeepMind	B-Application
submitted	O
its	O
initial	O
findings	O
in	O
a	O
paper	O
to	O
Nature	O
in	O
April	O
2017	O
,	O
which	O
was	O
then	O
published	O
in	O
October	O
2017	O
.	O
</s>
<s>
The	O
hardware	O
cost	O
for	O
a	O
single	O
AlphaGo	B-Application
Zero	I-Application
system	O
in	O
2017	O
,	O
including	O
the	O
four	O
TPUs	B-Device
,	O
has	O
been	O
quoted	O
as	O
around	O
$25	O
million	O
.	O
</s>
<s>
According	O
to	O
Hassabis	O
,	O
AlphaGo	B-Application
's	O
algorithms	O
are	O
likely	O
to	O
be	O
of	O
the	O
most	O
benefit	O
to	O
domains	O
that	O
require	O
an	O
intelligent	O
search	O
through	O
an	O
enormous	O
space	O
of	O
possibilities	O
,	O
such	O
as	O
protein	O
folding	O
(	O
see	O
AlphaFold	B-Application
)	O
or	O
accurately	O
simulating	O
chemical	O
reactions	O
.	O
</s>
<s>
AlphaGo	B-Application
's	O
techniques	O
are	O
probably	O
less	O
useful	O
in	O
domains	O
that	O
are	O
difficult	O
to	O
simulate	O
,	O
such	O
as	O
learning	O
how	O
to	O
drive	O
a	O
car	O
.	O
</s>
<s>
DeepMind	B-Application
stated	O
in	O
October	O
2017	O
that	O
it	O
had	O
already	O
started	O
active	O
work	O
on	O
attempting	O
to	O
use	O
AlphaGo	B-Application
Zero	I-Application
technology	O
for	O
protein	O
folding	O
,	O
and	O
stated	O
it	O
would	O
soon	O
publish	O
new	O
findings	O
.	O
</s>
<s>
AlphaGo	B-Application
Zero	I-Application
was	O
widely	O
regarded	O
as	O
a	O
significant	O
advance	O
,	O
even	O
when	O
compared	O
with	O
its	O
groundbreaking	O
predecessor	O
,	O
AlphaGo	B-Application
.	O
</s>
<s>
Oren	O
Etzioni	O
of	O
the	O
Allen	O
Institute	O
for	O
Artificial	B-Application
Intelligence	I-Application
called	O
AlphaGo	B-Application
Zero	I-Application
"	O
a	O
very	O
impressive	O
technical	O
result	O
"	O
in	O
"	O
both	O
their	O
ability	O
to	O
do	O
it	O
—	O
and	O
their	O
ability	O
to	O
train	O
the	O
system	O
in	O
40	O
days	O
,	O
on	O
four	O
TPUs	B-Device
"	O
.	O
</s>
<s>
The	O
Guardian	O
called	O
it	O
a	O
"	O
major	O
breakthrough	O
for	O
artificial	B-Application
intelligence	I-Application
"	O
,	O
citing	O
Eleni	O
Vasilaki	O
of	O
Sheffield	O
University	O
and	O
Tom	O
Mitchell	O
of	O
Carnegie	O
Mellon	O
University	O
,	O
who	O
called	O
it	O
an	O
impressive	O
feat	O
and	O
an	O
“	O
outstanding	O
engineering	O
accomplishment	O
"	O
respectively	O
.	O
</s>
<s>
Mark	O
Pesce	O
of	O
the	O
University	O
of	O
Sydney	O
called	O
AlphaGo	B-Application
Zero	I-Application
"	O
a	O
big	O
technological	O
advance	O
"	O
taking	O
us	O
into	O
"	O
undiscovered	O
territory	O
"	O
.	O
</s>
<s>
Gary	O
Marcus	O
,	O
a	O
psychologist	O
at	O
New	O
York	O
University	O
,	O
has	O
cautioned	O
that	O
for	O
all	O
we	O
know	O
,	O
AlphaGo	B-Application
may	O
contain	O
"	O
implicit	O
knowledge	O
that	O
the	O
programmers	O
have	O
about	O
how	O
to	O
construct	O
machines	O
to	O
play	O
problems	O
like	O
Go	O
"	O
and	O
will	O
need	O
to	O
be	O
tested	O
in	O
other	O
domains	O
before	O
being	O
sure	O
that	O
its	O
base	O
architecture	O
is	O
effective	O
at	O
much	O
more	O
than	O
playing	O
Go	O
.	O
</s>
<s>
In	O
contrast	O
,	O
DeepMind	B-Application
is	O
"	O
confident	O
that	O
this	O
approach	O
is	O
generalisable	O
to	O
a	O
large	O
number	O
of	O
domains	O
"	O
.	O
</s>
<s>
In	O
response	O
to	O
the	O
reports	O
,	O
South	O
Korean	O
Go	O
professional	O
Lee	O
Sedol	O
said	O
,	O
"	O
The	O
previous	O
version	O
of	O
AlphaGo	B-Application
was	O
n’t	O
perfect	O
,	O
and	O
I	O
believe	O
that	O
’s	O
why	O
AlphaGo	B-Application
Zero	I-Application
was	O
made.	O
"	O
</s>
<s>
On	O
the	O
potential	O
for	O
AlphaGo	B-Application
's	O
development	O
,	O
Lee	O
said	O
he	O
will	O
have	O
to	O
wait	O
and	O
see	O
but	O
also	O
said	O
it	O
will	O
affect	O
young	O
Go	O
players	O
.	O
</s>
<s>
Mok	O
Jin-seok	O
,	O
who	O
directs	O
the	O
South	O
Korean	O
national	O
Go	O
team	O
,	O
said	O
the	O
Go	O
world	O
has	O
already	O
been	O
imitating	O
the	O
playing	O
styles	O
of	O
previous	O
versions	O
of	O
AlphaGo	B-Application
and	O
creating	O
new	O
ideas	O
from	O
them	O
,	O
and	O
he	O
is	O
hopeful	O
that	O
new	O
ideas	O
will	O
come	O
out	O
from	O
AlphaGo	B-Application
Zero	I-Application
.	O
</s>
<s>
Mok	O
also	O
added	O
that	O
general	O
trends	O
in	O
the	O
Go	O
world	O
are	O
now	O
being	O
influenced	O
by	O
AlphaGo	B-Application
's	O
playing	O
style	O
.	O
</s>
<s>
"	O
We	O
are	O
now	O
past	O
the	O
point	O
where	O
we	O
debate	O
the	O
gap	O
between	O
the	O
capability	O
of	O
AlphaGo	B-Application
and	O
humans	O
.	O
</s>
<s>
Mok	O
has	O
reportedly	O
already	O
begun	O
analyzing	O
the	O
playing	O
style	O
of	O
AlphaGo	B-Application
Zero	I-Application
along	O
with	O
players	O
from	O
the	O
national	O
team	O
.	O
</s>
<s>
"	O
Though	O
having	O
watched	O
only	O
a	O
few	O
matches	O
,	O
we	O
received	O
the	O
impression	O
that	O
AlphaGo	B-Application
Zero	I-Application
plays	O
more	O
like	O
a	O
human	O
than	O
its	O
predecessors	O
,	O
"	O
Mok	O
said	O
.	O
</s>
<s>
Chinese	O
Go	O
professional	O
,	O
Ke	O
Jie	O
commented	O
on	O
the	O
remarkable	O
accomplishments	O
of	O
the	O
new	O
program	O
:	O
"	O
A	O
pure	O
self-learning	O
AlphaGo	B-Application
is	O
the	O
strongest	O
.	O
</s>
<s>
+Configuration	O
and	O
strengthVersionsPlaying	O
hardwareHardware	O
used	O
during	O
training	O
may	O
be	O
substantially	O
more	O
powerfulElo	O
ratingMatchesAlphaGo	O
Fan176	O
GPUs	B-Architecture
,	O
distributed3	O
,	O
1445:0	O
against	O
Fan	O
HuiAlphaGo	O
Lee48	O
TPUs	B-Device
,	O
distributed3	O
,	O
7394:1	O
against	O
Lee	O
SedolAlphaGo	O
Master4	O
TPUs	B-Device
,	O
single	O
machine4	O
,	O
85860:0	O
against	O
professional	O
players	O
;	O
</s>
<s>
89:11	O
against	O
AlphaGo	B-Application
MasterAlphaZero	O
4	O
TPUs	B-Device
,	O
single	O
machine	O
4,430	O
(	O
est	O
.	O
)	O
</s>
<s>
On	O
5	O
December	O
2017	O
,	O
DeepMind	B-Application
team	O
released	O
a	O
preprint	O
on	O
arXiv	O
,	O
introducing	O
AlphaZero	B-Application
,	O
a	O
program	O
using	O
generalized	O
AlphaGo	B-Application
Zero	I-Application
's	O
approach	O
,	O
which	O
achieved	O
within	O
24	O
hours	O
a	O
superhuman	O
level	O
of	O
play	O
in	O
chess	B-Application
,	O
shogi	O
,	O
and	O
Go	O
,	O
defeating	O
world-champion	O
programs	O
,	O
Stockfish	B-Application
,	O
Elmo	B-Application
,	O
and	O
3-day	O
version	O
of	O
AlphaGo	B-Application
Zero	I-Application
in	O
each	O
case	O
.	O
</s>
<s>
AlphaZero	B-Application
(	O
AZ	O
)	O
is	O
a	O
more	O
generalized	O
variant	O
of	O
the	O
AlphaGo	B-Application
Zero	I-Application
(	O
AGZ	O
)	O
algorithm	O
,	O
and	O
is	O
able	O
to	O
play	O
shogi	O
and	O
chess	B-Application
as	O
well	O
as	O
Go	O
.	O
</s>
<s>
AZ	O
has	O
hard-coded	O
rules	O
for	O
setting	O
search	O
hyperparameters	B-General_Concept
.	O
</s>
<s>
The	O
neural	B-Architecture
network	I-Architecture
is	O
now	O
updated	O
continually	O
.	O
</s>
<s>
Chess	B-Application
(	O
unlike	O
Go	O
)	O
can	O
end	O
in	O
a	O
tie	O
;	O
therefore	O
AZ	O
can	O
take	O
into	O
account	O
the	O
possibility	O
of	O
a	O
tie	O
game	O
.	O
</s>
<s>
An	O
open	B-License
source	I-License
program	O
,	O
Leela	B-Application
Zero	I-Application
,	O
based	O
on	O
the	O
ideas	O
from	O
the	O
AlphaGo	B-Application
papers	O
is	O
available	O
.	O
</s>
<s>
It	O
uses	O
a	O
GPU	B-Architecture
instead	O
of	O
the	O
TPUs	B-Device
recent	O
versions	O
of	O
AlphaGo	B-Application
rely	O
on	O
.	O
</s>
