<s>
Context	B-General_Concept
mixing	I-General_Concept
is	O
a	O
type	O
of	O
data	B-General_Concept
compression	I-General_Concept
algorithm	I-General_Concept
in	O
which	O
the	O
next-symbol	O
predictions	O
of	O
two	O
or	O
more	O
statistical	O
models	O
are	O
combined	O
to	O
yield	O
a	O
prediction	O
that	O
is	O
often	O
more	O
accurate	O
than	O
any	O
of	O
the	O
individual	O
predictions	O
.	O
</s>
<s>
The	O
random	B-Algorithm
forest	I-Algorithm
is	O
another	O
method	O
:	O
it	O
outputs	O
the	O
prediction	O
that	O
is	O
the	O
mode	O
of	O
the	O
predictions	O
output	O
by	O
individual	O
models	O
.	O
</s>
<s>
The	O
PAQ	B-Application
series	O
of	O
data	B-General_Concept
compression	I-General_Concept
programs	O
use	O
context	B-General_Concept
mixing	I-General_Concept
to	O
assign	O
probabilities	O
to	O
individual	O
bits	O
of	O
the	O
input	O
.	O
</s>
<s>
The	O
problem	O
is	O
important	O
for	O
data	B-General_Concept
compression	I-General_Concept
.	O
</s>
<s>
The	O
compression	B-General_Concept
ratio	I-General_Concept
depends	O
on	O
how	O
closely	O
the	O
estimated	O
probability	O
approaches	O
the	O
true	O
but	O
unknown	O
probability	O
of	O
event	O
.	O
</s>
<s>
Older	O
versions	O
of	O
PAQ	B-Application
uses	O
this	O
approach	O
.	O
</s>
<s>
Newer	O
versions	O
use	O
logistic	O
(	O
or	O
neural	B-Architecture
network	I-Architecture
)	O
mixing	O
by	O
first	O
transforming	O
the	O
predictions	O
into	O
the	O
logistic	O
domain	O
,	O
log( 	O
p/	O
(	O
1-p	O
)	O
)	O
before	O
averaging	O
.	O
</s>
<s>
All	O
but	O
the	O
oldest	O
versions	O
of	O
PAQ	B-Application
use	O
adaptive	O
weighting	O
.	O
</s>
<s>
Most	O
context	B-General_Concept
mixing	I-General_Concept
compressors	O
predict	O
one	O
bit	O
of	O
input	O
at	O
a	O
time	O
.	O
</s>
<s>
In	O
PAQ6	B-Application
,	O
whenever	O
one	O
of	O
the	O
bit	O
counts	O
is	O
incremented	O
,	O
the	O
part	O
of	O
the	O
other	O
count	O
that	O
exceeds	O
2	O
is	O
halved	O
.	O
</s>
<s>
All	O
PAQ	B-Application
versions	O
(	O
Matt	O
Mahoney	O
,	O
Serge	O
Osnach	O
,	O
Alexander	O
Ratushnyak	O
,	O
Przemysław	O
Skibiński	O
,	O
Jan	O
Ondrus	O
,	O
and	O
others	O
)	O
.	O
</s>
<s>
ZPAQ	B-Application
(	O
Matt	O
Mahoney	O
)	O
.	O
</s>
<s>
M1	O
and	O
M1X2	O
use	O
a	O
genetic	B-Algorithm
algorithm	I-Algorithm
to	O
select	O
two	O
bit	O
masked	O
contexts	O
in	O
a	O
separate	O
optimization	O
pass	O
.	O
</s>
<s>
enc	O
(	O
Serge	O
Osnach	O
)	O
tries	O
several	O
methods	O
based	O
on	O
PPM	B-Algorithm
and	O
(	O
linear	O
)	O
context	B-General_Concept
mixing	I-General_Concept
and	O
chooses	O
the	O
best	O
one	O
.	O
</s>
<s>
(	O
Byron	O
Knoll	O
)	O
mixes	O
many	O
models	O
,	O
and	O
is	O
currently	O
ranked	O
first	O
in	O
the	O
Large	O
Text	B-General_Concept
Compression	I-General_Concept
benchmark	O
,	O
as	O
well	O
as	O
the	O
Silesia	O
corpus	O
and	O
has	O
surpassed	O
the	O
winning	O
entry	O
of	O
the	O
Hutter	O
Prize	O
although	O
it	O
is	O
not	O
eligible	O
due	O
to	O
using	O
too	O
much	O
memory	O
.	O
</s>
