<s>
Leela	B-Application
Chess	I-Application
Zero	I-Application
(	O
abbreviated	O
as	O
LCZero	B-Application
,	O
lc0	B-Application
)	O
is	O
a	O
free	B-License
,	I-License
open-source	I-License
,	O
and	O
deep	O
neural	O
network	O
–	O
based	O
chess	B-Application
engine	I-Application
and	O
volunteer	B-Operating_System
computing	I-Operating_System
project	O
.	O
</s>
<s>
Development	O
has	O
been	O
spearheaded	O
by	O
programmer	O
Gary	O
Linscott	O
,	O
who	O
is	O
also	O
a	O
developer	O
for	O
the	O
Stockfish	B-Application
chess	I-Application
engine	I-Application
.	O
</s>
<s>
Leela	B-Application
Chess	I-Application
Zero	I-Application
was	O
adapted	O
from	O
the	O
Leela	B-Application
Zero	I-Application
Go	O
engine	O
,	O
which	O
in	O
turn	O
was	O
based	O
on	O
Google	B-Application
's	I-Application
AlphaGo	B-Application
Zero	I-Application
project	O
.	O
</s>
<s>
One	O
of	O
the	O
purposes	O
of	O
Leela	B-Application
Chess	I-Application
Zero	I-Application
was	O
to	O
verify	O
the	O
methods	O
in	O
the	O
AlphaZero	B-Application
paper	O
as	O
applied	O
to	O
the	O
game	O
of	O
chess	O
.	O
</s>
<s>
Like	O
Leela	B-Application
Zero	I-Application
and	O
AlphaGo	B-Application
Zero	I-Application
,	O
Leela	B-Application
Chess	I-Application
Zero	I-Application
starts	O
with	O
no	O
intrinsic	O
chess-specific	O
knowledge	O
other	O
than	O
the	O
basic	O
rules	O
of	O
the	O
game	O
.	O
</s>
<s>
Leela	B-Application
Chess	I-Application
Zero	I-Application
then	O
learns	O
how	O
to	O
play	O
chess	O
by	O
reinforcement	O
learning	O
from	O
repeated	O
self-play	B-Algorithm
,	O
using	O
a	O
distributed	O
computing	O
network	O
coordinated	O
at	O
the	O
Leela	B-Application
Chess	I-Application
Zero	I-Application
website	O
.	O
</s>
<s>
As	O
of	O
December	O
2022	O
,	O
Leela	B-Application
Chess	I-Application
Zero	I-Application
has	O
played	O
over	O
1.5	O
billion	O
games	O
against	O
itself	O
,	O
playing	O
around	O
1	O
million	O
games	O
every	O
day	O
,	O
and	O
is	O
capable	O
of	O
play	O
at	O
a	O
level	O
that	O
is	O
comparable	O
with	O
Stockfish	B-Application
,	O
the	O
leading	O
conventional	O
chess	B-Application
program	I-Application
.	O
</s>
<s>
The	O
Leela	B-Application
Chess	I-Application
Zero	I-Application
project	O
was	O
first	O
announced	O
on	O
TalkChess.com	O
on	O
January	O
9	O
,	O
2018	O
.	O
</s>
<s>
This	O
revealed	O
Leela	B-Application
Chess	I-Application
Zero	I-Application
as	O
the	O
open-source	O
,	O
self-learning	O
chess	B-Application
engine	I-Application
it	O
would	O
come	O
to	O
be	O
known	O
as	O
,	O
with	O
a	O
goal	O
of	O
creating	O
a	O
strong	O
chess	B-Application
engine	I-Application
.	O
</s>
<s>
Within	O
the	O
first	O
few	O
months	O
of	O
training	O
,	O
Leela	B-Application
Chess	I-Application
Zero	I-Application
had	O
already	O
reached	O
the	O
Grandmaster	O
level	O
,	O
surpassing	O
the	O
strength	O
of	O
early	O
releases	O
of	O
Rybka	B-Application
,	O
Stockfish	B-Application
,	O
and	O
Komodo	B-Application
,	O
despite	O
evaluating	O
orders	O
of	O
magnitude	O
fewer	O
positions	O
due	O
to	O
its	O
deep	O
neural	O
network	O
in	O
its	O
evaluation	B-General_Concept
function	I-General_Concept
and	O
its	O
use	O
of	O
Monte	B-Application
Carlo	I-Application
tree	I-Application
search	I-Application
.	O
</s>
<s>
In	O
December	O
2018	O
,	O
the	O
AlphaZero	B-Application
team	O
published	O
a	O
new	O
paper	O
in	O
Science	O
magazine	O
revealing	O
previously	O
undisclosed	O
details	O
of	O
the	O
architecture	O
and	O
training	O
parameters	O
used	O
for	O
AlphaZero	B-Application
.	O
</s>
<s>
These	O
changes	O
were	O
soon	O
incorporated	O
into	O
Leela	B-Application
Chess	I-Application
Zero	I-Application
and	O
increased	O
both	O
its	O
strength	O
and	O
training	O
efficiency	O
.	O
</s>
<s>
The	O
work	O
on	O
Leela	B-Application
Chess	I-Application
Zero	I-Application
has	O
informed	O
the	O
similar	O
AobaZero	O
project	O
for	O
shogi	O
.	O
</s>
<s>
The	O
engine	O
has	O
been	O
rewritten	O
and	O
carefully	O
iterated	O
upon	O
since	O
its	O
inception	O
,	O
and	O
now	O
runs	O
on	O
multiple	O
backends	B-Architecture
,	O
allowing	O
it	O
to	O
effectively	O
utilize	O
different	O
types	O
of	O
hardware	O
,	O
both	O
CPU	O
and	O
GPU	O
.	O
</s>
<s>
The	O
method	O
used	O
by	O
its	O
designers	O
to	O
make	O
Leela	B-Application
Chess	I-Application
Zero	I-Application
self-learn	O
and	O
play	O
chess	O
at	O
above	O
human	O
level	O
is	O
reinforcement	O
learning	O
.	O
</s>
<s>
This	O
is	O
a	O
machine-learning	O
algorithm	O
,	O
mirrored	O
from	O
AlphaZero	B-Application
used	O
by	O
the	O
Leela	B-Application
Chess	I-Application
Zero	I-Application
training	O
executable/binary	O
code	O
(	O
called	O
"	O
binary	O
"	O
)	O
to	O
maximize	O
reward	O
through	O
self-play	B-Algorithm
.	O
</s>
<s>
As	O
an	O
open-source	O
distributed	O
computing	O
project	O
,	O
volunteer	O
users	O
run	O
Leela	B-Application
Chess	I-Application
Zero	I-Application
to	O
play	O
hundreds	O
of	O
millions	O
of	O
games	O
which	O
are	O
fed	O
to	O
the	O
reinforcement	O
algorithm	O
.	O
</s>
<s>
In	O
order	O
to	O
contribute	O
to	O
the	O
advancement	O
of	O
the	O
Leela	B-Application
Chess	I-Application
Zero	I-Application
engine	O
,	O
the	O
latest	O
non-release	O
candidate	O
(	O
non-rc	O
)	O
version	O
of	O
the	O
Engine	O
as	O
well	O
as	O
the	O
Client	O
must	O
be	O
downloaded	O
.	O
</s>
<s>
The	O
Client	O
is	O
needed	O
to	O
connect	O
to	O
the	O
current	O
server	O
of	O
Leela	B-Application
Chess	I-Application
Zero	I-Application
,	O
where	O
all	O
of	O
the	O
information	O
from	O
the	O
self-play	B-Algorithm
chess	O
games	O
are	O
stored	O
,	O
to	O
obtain	O
the	O
latest	O
network	O
,	O
generate	O
self-play	B-Algorithm
games	O
,	O
and	O
upload	O
the	O
training	O
data	O
back	O
to	O
the	O
server	O
.	O
</s>
<s>
In	O
order	O
to	O
play	O
against	O
the	O
Leela	B-Application
Chess	I-Application
Zero	I-Application
engine	O
on	O
a	O
machine	O
,	O
two	O
components	O
are	O
needed	O
:	O
the	O
engine	O
binary	O
and	O
a	O
network	O
.	O
</s>
<s>
The	O
network	O
contains	O
Leela	B-Application
Chess	I-Application
Zero	I-Application
's	O
evaluation	B-General_Concept
function	I-General_Concept
that	O
is	O
needed	O
to	O
evaluate	O
positions	O
.	O
</s>
<s>
Older	O
networks	O
can	O
also	O
be	O
downloaded	O
and	O
used	O
by	O
placing	O
those	O
networks	O
in	O
the	O
folder	O
with	O
the	O
Lc0	B-Application
binary	O
.	O
</s>
<s>
Self-play	B-Algorithm
Elo	O
is	O
used	O
to	O
gauge	O
relative	O
network	O
strength	O
to	O
look	O
for	O
anomalies	O
and	O
general	O
changes	O
in	O
network	O
strength	O
,	O
and	O
can	O
be	O
used	O
as	O
a	O
diagnostic	O
tool	O
when	O
Lc0	B-Application
undergoes	O
significant	O
changes	O
.	O
</s>
<s>
Through	O
test	O
match	O
games	O
that	O
are	O
played	O
with	O
minimal	O
temperature-based	O
variation	O
,	O
Lc0	B-Application
engine	O
clients	O
test	O
the	O
most	O
recent	O
version	O
against	O
other	O
recent	O
versions	O
of	O
the	O
same	O
network	O
's	O
run	O
,	O
which	O
is	O
then	O
sent	O
to	O
the	O
training	O
server	O
to	O
create	O
an	O
overall	O
Elo	O
assessment	O
.	O
</s>
<s>
More	O
recent	O
Self-play	B-Algorithm
Elo	O
calculations	O
use	O
match	O
game	O
results	O
against	O
multiple	O
network	O
versions	O
to	O
calculate	O
a	O
more	O
accurate	O
Elo	O
value	O
.	O
</s>
<s>
The	O
Self-play	B-Algorithm
approach	O
has	O
several	O
consequences	O
on	O
gauging	O
Lc0	B-Application
Elo	O
rating	O
:	O
</s>
<s>
Measuring	O
Elo	O
relative	O
to	O
previous	O
networks	O
fails	O
to	O
measure	O
general	O
strength	O
since	O
networks	O
are	O
trained	O
to	O
anticipate	O
and	O
beat	O
the	O
predictions	O
made	O
by	O
prior	O
Lc0	B-Application
networks	O
rather	O
than	O
opponents	O
outside	O
the	O
training	O
domain	O
.	O
</s>
<s>
This	O
is	O
a	O
type	O
of	O
overfitting	B-Error_Name
measured	O
most	O
drastically	O
when	O
testing	O
smaller	O
networks	O
.	O
</s>
<s>
There	O
is	O
no	O
direct	O
one-to-one	O
correlation	O
between	O
self-play	B-Algorithm
Elo	O
and	O
the	O
strength	O
against	O
Alpha-Beta	O
engines	O
,	O
and	O
no	O
known	O
correlation	O
to	O
strength	O
against	O
humans	O
.	O
</s>
<s>
Cumulative	O
Self-Play	B-Algorithm
Elo	O
does	O
not	O
have	O
a	O
universal	O
conversion	O
to	O
conventional	O
Human	O
Elo	O
due	O
to	O
inflation	O
issues	O
presented	O
by	O
adversarial	O
play	O
,	O
and	O
the	O
reliance	O
of	O
the	O
measure	O
on	O
time	O
control	O
.	O
</s>
<s>
Cumulative	O
Self-Play	B-Algorithm
Elo	O
inflation	O
can	O
be	O
compared	O
with	O
other	O
runs	O
to	O
gauge	O
the	O
lack	O
of	O
generality	O
in	O
gauging	O
strength	O
with	O
pure	O
cumulative	O
self-play	B-Algorithm
elo	O
.	O
</s>
<s>
The	O
Fischer	O
Random	O
Chess	O
run	O
Test	O
71.4	O
(	O
named	O
714xxx	O
nets	O
)	O
,	O
ranks	O
at	O
nearly	O
4000	O
cumulative	O
self-play	B-Algorithm
Elo	O
76	O
nets	O
into	O
its	O
run	O
(	O
714076	O
)	O
.	O
</s>
<s>
The	O
T60	O
(	O
6xxxx	O
)	O
run	O
63000	O
net	O
has	O
a	O
cumulative	O
self-play	B-Algorithm
Elo	O
of	O
around	O
2900	O
.	O
</s>
<s>
This	O
alone	O
is	O
enough	O
to	O
credit	O
the	O
claim	O
that	O
Cumulative	O
self-play	B-Algorithm
Elo	O
is	O
not	O
an	O
objective	O
measure	O
of	O
strength	O
,	O
nor	O
is	O
it	O
a	O
measure	O
which	O
allows	O
one	O
to	O
linearly	O
compare	O
Lc0	B-Application
network	O
strength	O
to	O
Human	O
strength	O
.	O
</s>
<s>
Setting	O
up	O
the	O
engine	O
to	O
play	O
a	O
single	O
node	O
with	O
--minibatch-size	O
=	O
1	O
and	O
go	O
nodes	O
1	O
for	O
each	O
played	O
move	O
creates	O
deterministic	O
play	O
,	O
and	O
Self-Play	B-Algorithm
Elo	O
on	O
such	O
settings	O
will	O
always	O
yield	O
the	O
same	O
result	O
between	O
2	O
of	O
the	O
same	O
networks	O
on	O
the	O
same	O
start	O
position	O
—	O
always	O
win	O
,	O
always	O
loss	O
,	O
or	O
always	O
draw	O
.	O
</s>
<s>
Self-play	B-Algorithm
Elo	O
is	O
not	O
reliable	O
for	O
determining	O
strength	O
in	O
these	O
deterministic	O
circumstances	O
.	O
</s>
<s>
In	O
season	O
15	O
of	O
the	O
Top	O
Chess	B-Application
Engine	I-Application
Championship	O
,	O
the	O
engine	O
AllieStein	O
competed	O
alongside	O
Leela	O
.	O
</s>
<s>
AllieStein	O
is	O
a	O
combination	O
of	O
two	O
different	O
spinoffs	O
from	O
Leela	O
:	O
Allie	O
,	O
which	O
uses	O
the	O
same	O
evaluation	O
network	O
as	O
Leela	O
,	O
but	O
has	O
a	O
unique	O
search	O
algorithm	O
for	O
exploring	O
different	O
lines	O
of	O
play	O
,	O
and	O
Stein	O
,	O
an	O
evaluation	O
network	O
which	O
has	O
been	O
trained	O
using	O
supervised	B-General_Concept
learning	I-General_Concept
based	O
on	O
existing	O
game	O
data	O
featuring	O
other	O
engines	O
(	O
as	O
opposed	O
to	O
the	O
unsupervised	B-General_Concept
learning	I-General_Concept
which	O
Leela	O
uses	O
)	O
.	O
</s>
<s>
While	O
neither	O
of	O
these	O
projects	O
would	O
be	O
admitted	O
to	O
TCEC	O
separately	O
due	O
to	O
their	O
similarity	O
to	O
Leela	O
,	O
the	O
combination	O
of	O
Allie	O
's	O
search	O
algorithm	O
with	O
the	O
Stein	O
network	O
,	O
called	O
AllieStein	O
,	O
is	O
unique	O
enough	O
to	O
warrant	O
it	O
competing	O
alongside	O
mainstream	O
Lc0	B-Application
.	O
</s>
<s>
While	O
AllieStein	O
uses	O
the	O
same	O
code	O
to	O
evaluate	O
its	O
network	O
as	O
Lc0	B-Application
,	O
since	O
the	O
other	O
two	O
components	O
are	O
fresh	O
,	O
AllieStein	O
is	O
considered	O
a	O
distinct	O
engine	O
.	O
)	O
</s>
<s>
In	O
early	O
2021	O
,	O
the	O
LcZero	B-Application
blog	O
announced	O
Ceres	O
,	O
a	O
new	O
chess	B-Application
engine	I-Application
that	O
uses	O
LcZero	B-Application
networks	O
.	O
</s>
<s>
It	O
implements	O
Monte	B-Application
Carlo	I-Application
tree	I-Application
search	I-Application
as	O
well	O
as	O
many	O
novel	O
algorithmic	O
improvement	O
ideas	O
.	O
</s>
<s>
Initial	O
Elo	O
testing	O
showed	O
that	O
Ceres	O
is	O
stronger	O
than	O
Lc0	B-Application
with	O
the	O
same	O
network	O
.	O
</s>
<s>
In	O
April	O
2018	O
,	O
Leela	B-Application
Chess	I-Application
Zero	I-Application
became	O
the	O
first	O
engine	O
using	O
a	O
deep	O
neural	O
network	O
to	O
enter	O
the	O
Top	O
Chess	B-Application
Engine	I-Application
Championship	O
(	O
TCEC	O
)	O
,	O
during	O
season	O
12	O
in	O
the	O
lowest	O
division	O
,	O
division	O
4	O
.	O
</s>
<s>
Leela	O
then	O
won	O
the	O
30-game	O
match	O
against	O
Komodo	B-Application
to	O
secure	O
third	O
place	O
in	O
the	O
tournament	O
.	O
</s>
<s>
Leela	O
defeated	O
higher-division	O
engines	O
Laser	O
,	O
Ethereal	O
and	O
Fire	O
before	O
finally	O
being	O
eliminated	O
by	O
Stockfish	B-Application
in	O
the	O
semi-finals	O
.	O
</s>
<s>
Leela	O
finished	O
third	O
behind	O
Stockfish	B-Application
and	O
Komodo	B-Application
.	O
</s>
<s>
In	O
December	O
2018	O
,	O
Leela	O
participated	O
in	O
season	O
14	O
of	O
the	O
Top	O
Chess	B-Application
Engine	I-Application
Championship	O
.	O
</s>
<s>
In	O
the	O
premier	O
division	O
,	O
Stockfish	B-Application
dominated	O
while	O
Houdini	B-Application
,	O
Komodo	B-Application
and	O
Leela	O
competed	O
for	O
second	O
place	O
.	O
</s>
<s>
It	O
came	O
down	O
to	O
a	O
final-round	O
game	O
where	O
Leela	O
needed	O
to	O
hold	O
Stockfish	B-Application
to	O
a	O
draw	O
with	O
black	O
to	O
finish	O
second	O
ahead	O
of	O
Komodo	B-Application
.	O
</s>
<s>
Leela	O
managed	O
this	O
and	O
therefore	O
met	O
Stockfish	B-Application
in	O
the	O
superfinal	O
.	O
</s>
<s>
In	O
a	O
back	O
and	O
forth	O
match	O
,	O
first	O
Stockfish	B-Application
and	O
then	O
Leela	O
took	O
three	O
game	O
leads	O
before	O
Stockfish	B-Application
won	O
by	O
the	O
narrowest	O
margin	O
of	O
50.5	O
–	O
49.5	O
.	O
</s>
<s>
In	O
February	O
2019	O
,	O
Leela	O
scored	O
its	O
first	O
major	O
tournament	O
win	O
when	O
it	O
defeated	O
Houdini	B-Application
in	O
the	O
final	O
of	O
the	O
second	O
TCEC	O
cup	O
.	O
</s>
<s>
In	O
the	O
season	O
15	O
of	O
the	O
Top	O
Chess	B-Application
Engine	I-Application
Championship	O
(	O
May	O
2019	O
)	O
,	O
Leela	O
defended	O
its	O
TCEC	O
cup	O
title	O
,	O
this	O
time	O
defeating	O
Stockfish	B-Application
in	O
the	O
final	O
5.5	O
–	O
4.5	O
( +2	O
=	O
7	O
−1	O
)	O
after	O
Stockfish	B-Application
blundered	O
a	O
seven-man	O
tablebase	O
draw	O
.	O
</s>
<s>
Leela	O
also	O
won	O
the	O
Superfinal	O
for	O
the	O
first	O
time	O
,	O
scoring	O
53.5	O
–	O
46.5	O
( +14	O
−7	O
=	O
79	O
)	O
versus	O
Stockfish	B-Application
,	O
including	O
winning	O
as	O
both	O
white	O
and	O
black	O
in	O
the	O
same	O
predetermined	O
opening	O
in	O
games	O
61	O
and	O
62	O
.	O
</s>
<s>
Season	O
16	O
of	O
TCEC	O
saw	O
Leela	O
finish	O
in	O
third	O
place	O
in	O
premier	O
division	O
,	O
missing	O
qualification	O
for	O
the	O
superfinal	O
to	O
Stockfish	B-Application
and	O
new	O
deep	O
neural	O
network	O
engine	O
AllieStein	O
.	O
</s>
<s>
Leela	O
did	O
not	O
suffer	O
any	O
losses	O
in	O
the	O
Premier	O
division	O
,	O
the	O
only	O
engine	O
to	O
do	O
so	O
,	O
and	O
defeated	O
Stockfish	B-Application
in	O
one	O
of	O
the	O
six	O
games	O
they	O
played	O
.	O
</s>
<s>
However	O
,	O
Leela	O
only	O
managed	O
to	O
score	O
nine	O
wins	O
,	O
while	O
AllieStein	O
and	O
Stockfish	B-Application
both	O
scored	O
14	O
wins	O
.	O
</s>
<s>
This	O
inability	O
to	O
defeat	O
weaker	O
engines	O
led	O
to	O
Leela	O
finishing	O
third	O
,	O
half	O
a	O
point	O
behind	O
AllieStein	O
and	O
a	O
point	O
behind	O
Stockfish	B-Application
.	O
</s>
<s>
In	O
the	O
fourth	O
TCEC	O
cup	O
,	O
Leela	O
was	O
seeded	O
first	O
as	O
the	O
defending	O
champion	O
,	O
which	O
placed	O
it	O
on	O
the	O
opposite	O
half	O
of	O
the	O
brackets	O
as	O
AllieStein	O
and	O
Stockfish	B-Application
.	O
</s>
<s>
Leela	O
was	O
able	O
to	O
qualify	O
for	O
the	O
finals	O
,	O
where	O
it	O
faced	O
Stockfish	B-Application
.	O
</s>
<s>
After	O
seven	O
draws	O
,	O
Stockfish	B-Application
won	O
the	O
eighth	O
game	O
to	O
win	O
the	O
match	O
.	O
</s>
<s>
In	O
Season	O
17	O
of	O
TCEC	O
,	O
held	O
in	O
January	O
–	O
April	O
2020	O
,	O
Leela	O
regained	O
the	O
championship	O
by	O
defeating	O
Stockfish	B-Application
52.5	O
–	O
47.5	O
,	O
scoring	O
a	O
remarkable	O
six	O
wins	O
in	O
the	O
final	O
ten	O
games	O
,	O
including	O
winning	O
as	O
both	O
white	O
and	O
black	O
in	O
the	O
same	O
predetermined	O
opening	O
in	O
games	O
95	O
and	O
96	O
.	O
</s>
<s>
It	O
qualified	O
for	O
the	O
superfinal	O
again	O
in	O
Season	O
18	O
,	O
but	O
this	O
time	O
was	O
defeated	O
by	O
Stockfish	B-Application
53.5	O
–	O
46.5	O
.	O
</s>
<s>
This	O
time	O
it	O
played	O
against	O
a	O
new	O
Stockfish	B-Application
version	O
with	O
support	O
for	O
NNUE	B-General_Concept
,	O
a	O
shallow	O
neural	O
network	O
–	O
based	O
evaluation	B-General_Concept
function	I-General_Concept
used	O
primarily	O
for	O
the	O
leaf	O
nodes	O
of	O
the	O
search	O
tree	O
.	O
</s>
<s>
Stockfish	B-Application
NNUE	B-General_Concept
defeated	O
Leela	O
convincingly	O
with	O
a	O
final	O
score	O
of	O
54.5	O
–	O
45.5	O
( +18	O
−9	O
=	O
73	O
)	O
.	O
</s>
<s>
Since	O
then	O
,	O
Leela	O
has	O
qualified	O
three	O
times	O
for	O
the	O
superfinal	O
,	O
only	O
to	O
lose	O
every	O
time	O
to	O
Stockfish	B-Application
:	O
+14	O
−8	O
=	O
78	O
in	O
Season	O
20	O
,	O
+19	O
−7	O
=	O
74	O
in	O
Season	O
21	O
and	O
+27	O
−10	O
=	O
63	O
in	O
Season	O
23	O
.	O
</s>
<s>
Leela	O
beats	O
the	O
world	O
champion	O
Stockfish	B-Application
engine	O
despite	O
a	O
one-pawn	O
handicap	O
.	O
</s>
<s>
Leela	O
completely	O
outplayed	O
Stockfish	B-Application
with	O
black	O
pieces	O
in	O
Trompovsky	O
attack	O
,	O
Leela	O
's	O
eval	O
went	O
from	O
0.1	O
to	O
−	O
1.2	O
in	O
one	O
move	O
,	O
and	O
Stockfish	B-Application
's	O
eval	O
did	O
not	O
go	O
negative	O
until	O
15	O
moves	O
later	O
.	O
</s>
