<s>
Rprop	B-Algorithm
,	O
short	O
for	O
resilient	B-Algorithm
backpropagation	I-Algorithm
,	O
is	O
a	O
learning	O
heuristic	B-Algorithm
for	O
supervised	B-General_Concept
learning	I-General_Concept
in	O
feedforward	B-Algorithm
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
.	O
</s>
<s>
This	O
is	O
a	O
first-order	B-Algorithm
optimization	O
algorithm	O
.	O
</s>
<s>
Similarly	O
to	O
the	O
Manhattan	O
update	O
rule	O
,	O
Rprop	B-Algorithm
takes	O
into	O
account	O
only	O
the	O
sign	O
of	O
the	O
partial	O
derivative	O
over	O
all	O
patterns	O
(	O
not	O
the	O
magnitude	O
)	O
,	O
and	O
acts	O
independently	O
on	O
each	O
"	O
weight	O
"	O
.	O
</s>
<s>
Rprop	B-Algorithm
can	O
result	O
in	O
very	O
large	O
weight	O
increments	O
or	O
decrements	O
if	O
the	O
gradients	O
are	O
large	O
,	O
which	O
is	O
a	O
problem	O
when	O
using	O
mini-batches	O
as	O
opposed	O
to	O
full	O
batches	O
.	O
</s>
<s>
RPROP	B-Algorithm
is	O
a	O
batch	O
update	O
algorithm	O
.	O
</s>
<s>
Next	O
to	O
the	O
cascade	O
correlation	O
algorithm	O
and	O
the	O
Levenberg	O
–	O
Marquardt	O
algorithm	O
,	O
Rprop	B-Algorithm
is	O
one	O
of	O
the	O
fastest	O
weight	O
update	O
mechanisms	O
.	O
</s>
<s>
Martin	O
Riedmiller	O
developed	O
three	O
algorithms	O
,	O
all	O
named	O
RPROP	B-Algorithm
.	O
</s>
<s>
RPROP+	O
is	O
defined	O
at	O
.	O
</s>
<s>
RPROP	B-Algorithm
is	O
defined	O
at	O
.	O
</s>
<s>
Backtracking	O
is	O
removed	O
from	O
RPROP+	O
.	O
</s>
