<s>
In	O
constraint	B-Application
satisfaction	I-Application
backtracking	B-Algorithm
algorithms	I-Algorithm
,	O
constraint	B-Application
learning	I-Application
is	O
a	O
technique	O
for	O
improving	O
efficiency	O
.	O
</s>
<s>
Clause	B-Application
learning	I-Application
is	O
the	O
name	O
of	O
this	O
technique	O
when	O
applied	O
to	O
propositional	B-Algorithm
satisfiability	I-Algorithm
.	O
</s>
<s>
Backtracking	B-Algorithm
algorithms	I-Algorithm
work	O
by	O
choosing	O
an	O
unassigned	O
variable	O
and	O
recursively	O
solve	O
the	O
problems	O
obtained	O
by	O
assigning	O
a	O
value	O
to	O
this	O
variable	O
.	O
</s>
<s>
A	O
constraint	B-Application
learning	I-Application
algorithm	O
differs	O
because	O
it	O
tries	O
to	O
record	O
some	O
information	O
,	O
before	O
backtracking	B-Algorithm
,	O
in	O
form	O
of	O
a	O
new	O
constraint	O
.	O
</s>
<s>
If	O
the	O
algorithm	O
has	O
learned	O
the	O
new	O
constraint	O
,	O
it	O
will	O
backtrack	O
from	O
this	O
solution	O
,	O
while	O
the	O
original	O
backtracking	B-Algorithm
algorithm	I-Algorithm
would	O
do	O
a	O
subsequent	O
search	O
.	O
</s>
<s>
However	O
,	O
recording	O
this	O
constraint	O
is	O
not	O
useful	O
,	O
as	O
this	O
partial	O
solution	O
will	O
not	O
be	O
encountered	O
again	O
due	O
to	O
the	O
way	O
backtracking	B-Algorithm
proceeds	O
.	O
</s>
<s>
The	O
efficiency	O
of	O
constraint	B-Application
learning	I-Application
algorithm	O
is	O
balanced	O
between	O
two	O
factors	O
.	O
</s>
<s>
On	O
one	O
hand	O
,	O
the	O
more	O
often	O
a	O
recorded	O
constraint	O
is	O
violated	O
,	O
the	O
more	O
often	O
backtracking	B-Algorithm
avoids	O
doing	O
useless	O
search	O
.	O
</s>
<s>
Various	O
constraint	B-Application
learning	I-Application
techniques	O
exist	O
,	O
differing	O
in	O
strictness	O
of	O
recorded	O
constraints	O
and	O
cost	O
of	O
finding	O
them	O
.	O
</s>
<s>
It	O
uses	O
the	O
same	O
rationale	O
of	O
graph-based	B-Application
backjumping	I-Application
.	O
</s>
<s>
These	O
methods	O
are	O
called	O
"	O
graph-based	O
"	O
because	O
they	O
are	O
based	O
on	O
pairs	O
of	O
variables	O
are	O
in	O
the	O
same	O
constraint	O
,	O
which	O
can	O
be	O
found	O
out	O
from	O
the	O
graph	O
associated	O
to	O
the	O
constraint	B-Application
satisfaction	I-Application
problem	I-Application
.	O
</s>
<s>
Jumpback	O
learning	O
is	O
based	O
on	O
storing	O
as	O
constraints	O
the	O
inconsistent	O
assignments	O
that	O
would	O
be	O
found	O
by	O
conflict-based	B-Application
backjumping	I-Application
.	O
</s>
<s>
Constraint	B-Application
learning	I-Application
algorithms	O
differ	O
not	O
only	O
on	O
the	O
choice	O
of	O
constraint	O
corresponding	O
to	O
a	O
given	O
inconsistent	O
partial	O
evaluation	O
,	O
but	O
also	O
on	O
the	O
choice	O
of	O
which	O
constraint	O
they	O
maintain	O
and	O
which	O
ones	O
they	O
discard	O
.	O
</s>
<s>
Relevance-bounded	B-Application
learning	I-Application
discards	O
constraints	O
(	O
or	O
does	O
not	O
store	O
them	O
at	O
all	O
)	O
that	O
are	O
considered	O
not	O
relevant	O
given	O
the	O
current	O
point	O
of	O
the	O
search	O
space	O
;	O
in	O
particular	O
,	O
it	O
discards	O
or	O
does	O
not	O
store	O
all	O
constraints	O
that	O
represent	O
inconsistent	O
partial	O
evaluations	O
that	O
differ	O
from	O
the	O
current	O
partial	O
evaluation	O
on	O
no	O
more	O
than	O
a	O
given	O
fixed	O
number	O
of	O
variables	O
.	O
</s>
