<s>
MuZero	B-Application
is	O
a	O
computer	B-Application
program	I-Application
developed	O
by	O
artificial	B-Application
intelligence	I-Application
research	I-Application
company	O
DeepMind	B-Application
to	O
master	O
games	O
without	O
knowing	O
their	O
rules	O
.	O
</s>
<s>
Its	O
release	O
in	O
2019	O
included	O
benchmarks	O
of	O
its	O
performance	O
in	O
go	O
,	O
chess	B-Application
,	O
shogi	O
,	O
and	O
a	O
standard	O
suite	O
of	O
Atari	O
games	O
.	O
</s>
<s>
The	O
algorithm	O
uses	O
an	O
approach	O
similar	O
to	O
AlphaZero	B-Application
.	O
</s>
<s>
It	O
matched	O
AlphaZero	B-Application
's	O
performance	O
in	O
chess	B-Application
and	O
shogi	O
,	O
improved	O
on	O
its	O
performance	O
in	O
Go	O
(	O
setting	O
a	O
new	O
world	O
record	O
)	O
,	O
and	O
improved	O
on	O
the	O
state	O
of	O
the	O
art	O
in	O
mastering	O
a	O
suite	O
of	O
57	O
Atari	O
games	O
(	O
the	O
Arcade	O
Learning	O
Environment	O
)	O
,	O
a	O
visually-complex	O
domain	O
.	O
</s>
<s>
MuZero	B-Application
was	O
trained	O
via	O
self-play	B-Algorithm
,	O
with	O
no	O
access	O
to	O
rules	O
,	O
opening	O
books	O
,	O
or	O
endgame	O
tablebases	O
.	O
</s>
<s>
The	O
trained	O
algorithm	O
used	O
the	O
same	O
convolutional	O
and	O
residual	O
algorithms	O
as	O
AlphaZero	B-Application
,	O
but	O
with	O
20	O
percent	O
fewer	O
computation	O
steps	O
per	O
node	O
in	O
the	O
search	O
tree	O
.	O
</s>
<s>
On	O
November	O
19	O
,	O
2019	O
,	O
the	O
DeepMind	B-Application
team	O
released	O
a	O
preprint	O
introducing	O
MuZero	B-Application
.	O
</s>
<s>
MuZero	B-Application
(	O
MZ	O
)	O
is	O
a	O
combination	O
of	O
the	O
high-performance	O
planning	O
of	O
the	O
AlphaZero	B-Application
(	O
AZ	O
)	O
algorithm	O
with	O
approaches	O
to	O
model-free	O
reinforcement	O
learning	O
.	O
</s>
<s>
MuZero	B-Application
was	O
derived	O
directly	O
from	O
AZ	O
code	O
,	O
sharing	O
its	O
rules	O
for	O
setting	O
hyperparameters	B-General_Concept
.	O
</s>
<s>
AZ	O
's	O
planning	O
process	O
uses	O
a	O
simulator	B-Application
.	O
</s>
<s>
The	O
simulator	B-Application
knows	O
the	O
rules	O
of	O
the	O
game	O
.	O
</s>
<s>
A	O
neural	B-Architecture
network	I-Architecture
then	O
predicts	O
the	O
policy	O
and	O
value	O
of	O
a	O
future	O
position	O
.	O
</s>
<s>
MZ	O
does	O
not	O
have	O
access	O
to	O
the	O
rules	O
,	O
and	O
instead	O
learns	O
one	O
with	O
neural	B-Architecture
networks	I-Architecture
.	O
</s>
<s>
MuZero	B-Application
surpassed	O
both	O
R2D2	O
's	O
mean	O
and	O
median	O
performance	O
across	O
the	O
suite	O
of	O
games	O
,	O
though	O
it	O
did	O
not	O
do	O
better	O
in	O
every	O
game	O
.	O
</s>
<s>
MuZero	B-Application
used	O
16	O
third-generation	O
tensor	B-Device
processing	I-Device
units	I-Device
(	O
TPUs	O
)	O
for	O
training	O
,	O
and	O
1000	O
TPUs	O
for	O
selfplay	O
for	O
board	O
games	O
,	O
with	O
800	O
simulations	O
per	O
step	O
and	O
8	O
TPUs	O
for	O
training	O
and	O
32	O
TPUs	O
for	O
selfplay	O
for	O
Atari	O
games	O
,	O
with	O
50	O
simulations	O
per	O
step	O
.	O
</s>
<s>
AlphaZero	B-Application
used	O
64	O
second-generation	O
TPUs	O
for	O
training	O
,	O
and	O
5000	O
first-generation	O
TPUs	O
for	O
selfplay	O
.	O
</s>
<s>
MuZero	B-Application
matched	O
AlphaZero	B-Application
's	O
performance	O
in	O
chess	B-Application
and	O
Shogi	O
after	O
roughly	O
1	O
million	O
training	O
steps	O
.	O
</s>
<s>
MuZero	B-Application
was	O
viewed	O
as	O
a	O
significant	O
advancement	O
over	O
AlphaZero	B-Application
,	O
and	O
a	O
generalizable	O
step	O
forward	O
in	O
unsupervised	B-General_Concept
learning	I-General_Concept
techniques	O
.	O
</s>
<s>
MuZero	B-Application
has	O
been	O
used	O
as	O
a	O
reference	O
implementation	O
in	O
other	O
work	O
,	O
for	O
instance	O
as	O
a	O
way	O
to	O
generate	O
model-based	O
behavior	O
.	O
</s>
<s>
In	O
late	O
2021	O
,	O
a	O
more	O
efficient	O
variant	O
of	O
MuZero	B-Application
was	O
proposed	O
,	O
named	O
EfficientZero	O
.	O
</s>
