<s>
In	O
machine	O
learning	O
,	O
particularly	O
in	O
the	O
creation	O
of	O
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
,	O
ensemble	B-General_Concept
averaging	I-General_Concept
is	O
the	O
process	O
of	O
creating	O
multiple	O
models	O
and	O
combining	O
them	O
to	O
produce	O
a	O
desired	O
output	O
,	O
as	O
opposed	O
to	O
creating	O
just	O
one	O
model	O
.	O
</s>
<s>
Ensemble	B-General_Concept
averaging	I-General_Concept
is	O
one	O
of	O
the	O
simplest	O
types	O
of	O
committee	B-Algorithm
machines	I-Algorithm
.	O
</s>
<s>
Along	O
with	O
boosting	B-Algorithm
,	O
it	O
is	O
one	O
of	O
the	O
two	O
major	O
types	O
of	O
static	O
committee	B-Algorithm
machines	I-Algorithm
.	O
</s>
<s>
In	O
contrast	O
to	O
standard	O
network	O
design	O
in	O
which	O
many	O
networks	O
are	O
generated	O
but	O
only	O
one	O
is	O
kept	O
,	O
ensemble	B-General_Concept
averaging	I-General_Concept
keeps	O
the	O
less	O
satisfactory	O
networks	O
around	O
,	O
but	O
with	O
less	O
weight	O
.	O
</s>
<s>
The	O
theory	O
of	O
ensemble	B-General_Concept
averaging	I-General_Concept
relies	O
on	O
two	O
properties	O
of	O
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
:	O
</s>
<s>
Ensemble	B-General_Concept
averaging	I-General_Concept
creates	O
a	O
group	O
of	O
networks	O
,	O
each	O
with	O
low	O
bias	O
and	O
high	O
variance	O
,	O
then	O
combines	O
them	O
to	O
a	O
new	O
network	O
with	O
(	O
hopefully	O
)	O
low	O
bias	O
and	O
low	O
variance	O
.	O
</s>
<s>
It	O
is	O
thus	O
a	O
resolution	O
of	O
the	O
bias-variance	B-General_Concept
dilemma	I-General_Concept
.	O
</s>
<s>
A	O
more	O
complex	O
version	O
of	O
ensemble	B-General_Concept
average	I-General_Concept
views	O
the	O
final	O
result	O
not	O
as	O
a	O
mere	O
average	O
of	O
all	O
the	O
experts	O
,	O
but	O
rather	O
as	O
a	O
weighted	O
sum	O
.	O
</s>
<s>
The	O
optimization	O
problem	O
of	O
finding	O
alpha	O
is	O
readily	O
solved	O
through	O
neural	B-Architecture
networks	I-Architecture
,	O
hence	O
a	O
"	O
meta-network	O
"	O
where	O
each	O
"	O
neuron	O
"	O
is	O
in	O
fact	O
an	O
entire	O
neural	B-Architecture
network	I-Architecture
can	O
be	O
trained	O
,	O
and	O
the	O
synaptic	O
weights	O
of	O
the	O
final	O
network	O
is	O
the	O
weight	O
applied	O
to	O
each	O
expert	O
.	O
</s>
<s>
It	O
can	O
be	O
seen	O
that	O
most	O
forms	O
of	O
neural	B-Architecture
networks	I-Architecture
are	O
some	O
subset	O
of	O
a	O
linear	O
combination	O
:	O
the	O
standard	O
neural	B-Architecture
net	I-Architecture
(	O
where	O
only	O
one	O
expert	O
is	O
used	O
)	O
is	O
simply	O
a	O
linear	O
combination	O
with	O
all	O
and	O
one	O
.	O
</s>
<s>
A	O
more	O
recent	O
ensemble	B-General_Concept
averaging	I-General_Concept
method	O
is	O
negative	O
correlation	O
learning	O
,	O
proposed	O
by	O
Y	O
.	O
Liu	O
and	O
X	O
.	O
Yao	O
.	O
</s>
