<s>
In	O
computer	B-General_Concept
science	I-General_Concept
,	O
beam	B-Algorithm
search	I-Algorithm
is	O
a	O
heuristic	B-Algorithm
search	B-Application
algorithm	I-Application
that	O
explores	O
a	O
graph	O
by	O
expanding	O
the	O
most	O
promising	O
node	O
in	O
a	O
limited	O
set	O
.	O
</s>
<s>
Beam	B-Algorithm
search	I-Algorithm
is	O
an	O
optimization	O
of	O
best-first	B-Algorithm
search	I-Algorithm
that	O
reduces	O
its	O
memory	O
requirements	O
.	O
</s>
<s>
Best-first	B-Algorithm
search	I-Algorithm
is	O
a	O
graph	O
search	O
which	O
orders	O
all	O
partial	O
solutions	O
(	O
states	O
)	O
according	O
to	O
some	O
heuristic	B-Algorithm
.	O
</s>
<s>
But	O
in	O
beam	B-Algorithm
search	I-Algorithm
,	O
only	O
a	O
predetermined	O
number	O
of	O
best	O
partial	O
solutions	O
are	O
kept	O
as	O
candidates	O
.	O
</s>
<s>
It	O
is	O
thus	O
a	O
greedy	B-Algorithm
algorithm	I-Algorithm
.	O
</s>
<s>
The	O
term	O
"	O
beam	B-Algorithm
search	I-Algorithm
"	O
was	O
coined	O
by	O
Raj	O
Reddy	O
of	O
Carnegie	O
Mellon	O
University	O
in	O
1977	O
.	O
</s>
<s>
Beam	B-Algorithm
search	I-Algorithm
uses	O
breadth-first	B-Algorithm
search	I-Algorithm
to	O
build	O
its	O
search	B-Algorithm
tree	I-Algorithm
.	O
</s>
<s>
At	O
each	O
level	O
of	O
the	O
tree	O
,	O
it	O
generates	O
all	O
successors	O
of	O
the	O
states	O
at	O
the	O
current	O
level	O
,	O
sorting	O
them	O
in	O
increasing	O
order	O
of	O
heuristic	B-Algorithm
cost	O
.	O
</s>
<s>
With	O
an	O
infinite	O
beam	O
width	O
,	O
no	O
states	O
are	O
pruned	O
and	O
beam	B-Algorithm
search	I-Algorithm
is	O
identical	O
to	O
breadth-first	B-Algorithm
search	I-Algorithm
.	O
</s>
<s>
Since	O
a	O
goal	O
state	O
could	O
potentially	O
be	O
pruned	O
,	O
beam	B-Algorithm
search	I-Algorithm
sacrifices	O
completeness	O
(	O
the	O
guarantee	O
that	O
an	O
algorithm	O
will	O
terminate	O
with	O
a	O
solution	O
,	O
if	O
one	O
exists	O
)	O
.	O
</s>
<s>
Beam	B-Algorithm
search	I-Algorithm
is	O
not	O
optimal	O
(	O
that	O
is	O
,	O
there	O
is	O
no	O
guarantee	O
that	O
it	O
will	O
find	O
the	O
best	O
solution	O
)	O
.	O
</s>
<s>
A	O
beam	B-Algorithm
search	I-Algorithm
is	O
most	O
often	O
used	O
to	O
maintain	O
tractability	O
in	O
large	O
systems	O
with	O
insufficient	O
amount	O
of	O
memory	O
to	O
store	O
the	O
entire	O
search	B-Algorithm
tree	I-Algorithm
.	O
</s>
<s>
For	O
example	O
,	O
it	O
has	O
been	O
used	O
in	O
many	O
machine	B-Application
translation	I-Application
systems	I-Application
.	O
</s>
<s>
(	O
The	O
state	O
of	O
the	O
art	O
now	O
primarily	O
uses	O
neural	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
based	O
methods	O
.	O
)	O
</s>
<s>
The	O
first	O
use	O
of	O
a	O
beam	B-Algorithm
search	I-Algorithm
was	O
in	O
the	O
Harpy	O
Speech	O
Recognition	O
System	O
,	O
CMU	O
1976	O
.	O
</s>
<s>
Beam	B-Algorithm
search	I-Algorithm
has	O
been	O
made	O
complete	O
by	O
combining	O
it	O
with	O
depth-first	B-Algorithm
search	I-Algorithm
,	O
resulting	O
in	O
beam	B-Algorithm
stack	I-Algorithm
search	I-Algorithm
and	O
depth-first	B-Algorithm
beam	B-Algorithm
search	I-Algorithm
,	O
and	O
with	O
limited	O
discrepancy	O
search	O
,	O
resulting	O
in	O
beam	B-Algorithm
search	I-Algorithm
using	O
limited	O
discrepancy	O
backtracking	O
(	O
BULB	O
)	O
.	O
</s>
<s>
The	O
resulting	O
search	B-Application
algorithms	I-Application
are	O
anytime	B-Algorithm
algorithms	I-Algorithm
that	O
find	O
good	O
but	O
likely	O
sub-optimal	O
solutions	O
quickly	O
,	O
like	O
beam	B-Algorithm
search	I-Algorithm
,	O
then	O
backtrack	O
and	O
continue	O
to	O
find	O
improved	O
solutions	O
until	O
convergence	O
to	O
an	O
optimal	O
solution	O
.	O
</s>
<s>
In	O
the	O
context	O
of	O
a	O
local	B-Algorithm
search	I-Algorithm
,	O
we	O
call	O
local	O
beam	B-Algorithm
search	I-Algorithm
a	O
specific	O
algorithm	O
that	O
begins	O
selecting	O
randomly	O
generated	O
states	O
and	O
then	O
,	O
for	O
each	O
level	O
of	O
the	O
search	B-Algorithm
tree	I-Algorithm
,	O
it	O
always	O
considers	O
new	O
states	O
among	O
all	O
the	O
possible	O
successors	O
of	O
the	O
current	O
ones	O
,	O
until	O
it	O
reaches	O
a	O
goal	O
.	O
</s>
<s>
Since	O
local	O
beam	B-Algorithm
search	I-Algorithm
often	O
ends	O
up	O
on	O
local	O
maxima	O
,	O
a	O
common	O
solution	O
is	O
to	O
choose	O
the	O
next	O
states	O
in	O
a	O
random	O
way	O
,	O
with	O
a	O
probability	O
dependent	O
from	O
the	O
heuristic	B-Algorithm
evaluation	O
of	O
the	O
states	O
.	O
</s>
<s>
This	O
kind	O
of	O
search	O
is	O
called	O
stochastic	O
beam	B-Algorithm
search	I-Algorithm
.	O
</s>
<s>
Other	O
variants	O
are	O
flexible	O
beam	B-Algorithm
search	I-Algorithm
and	O
recovery	O
beam	B-Algorithm
search	I-Algorithm
.	O
</s>
